ASHS 2025 Conference

Luke Concollato will provide an introduction to new, intuitive platform for digital phenotyping and integration of environmental data. Luke Concollato will introduce the program and give instructions on how this platform can be used for specified organization of data and ultimately, breakthroughs in plant science.

This is a pre-recorded Zoom webinar that will be shown at the Hort Theater.

Vidalia sweet onion (Allium cepa spp.), holds significant economic and cultural value in the Vidalia region, known for its sweet flavor and low pungency. The success of marketable bulb production is closely linked to the early crop stages, particularly seedling production and field-quality transplanting. Due to the limited availability of mechanical transplanters, transplanting is predominantly performed manually, resulting in variability in planting quality. A major issue is the presence of seedlings lying on the ground, which significantly reduces survival rates and leads to undersized bulbs. Assessing the number of such seedlings manually is labor-intensive and inefficient. Therefore, this study aimed to develop a rapid and accurate method to quantify the number of seedlings lying on the ground across the field using a unmanned aerial vehicle (UAV) imagery and the YOLOv8 deep learning model. On January 6, 2025, aerial RGB images were captured using a UAV over an experimental onion field located at the UGA Vidalia Onion and Vegetable Research Center in Lyons, GA. The seedlings were transplanted on December 15, 2024, under three treatment conditions: 100% properly transplanted plants, 100% lying on the ground, and 25% lying on the ground. The objective of this study was to detect and quantify the percentage of plants lying on the ground within each plot using computer vision techniques. The images were manually annotated and used to train a YOLOv8 object detection model. The dataset was split into 70% for training, 15% for validation, and 15% for testing. Model performance was evaluated using standard YOLOv8 accuracy metrics, including precision, recall, and F1-score. The YOLOv8 model demonstrated strong performance in detecting onion seedlings lying on the ground. It achieved moderate to high precision and recall, as well as a high F1-score on the test dataset. When applied to the annotated aerial images, the model accurately quantified the percentage of fallen seedlings within each treatment. In the 100% properly transplanted treatment, a small number of seedlings were falsely identified as lying on the ground. These results confirm the model's ability to reliably distinguish between transplant quality levels using UAV imagery. This approach offers a scalable solution for monitoring transplant quality and identifying problematic planting zones.

This study developed an AI-powered image preprocessing tool to analyze daylily flower traits, providing a user-friendly and feature-rich solution for accurate image segmentation- a process that separates the flower from the background in photographs. The tool's design combines a pre-trained TRACER convolutional neural network (CNN) with an EfficientNet backbone to ensure robust and precise segmentation. It features a user-friendly graphical interface built with PyQt that facilitates seamless image uploads, real-time visualization of segmentation results, and customization options. Furthermore, a built-in manual refinement pencil function allows for precise corrections to automated masks, ensuring accuracy when needed. It supports multiple output formats, including masks and annotated images, enabling easy integration into subsequent analyses. A dataset comprising 1,450 daylily color images was employed for training and evaluating the tool. A comparative analysis of four deep learning models (TRACER, BASNet, U2 Net, and DeepLabV3), using metrics such as Accuracy, Precision, Recall, Dice Score, Jaccard Index, and XOR Error, highlighted the superior performance of the TRACER model in terms of accuracy and reliability. The findings show that the tool successfully separates flowers from intricate backgrounds, effectively addressing issues such as petal overlap, various phenotypes, and environmental variations. This automation tool has the potential to replace the labor-intensive and error-prone manual segmentation process, which has posed a significant limitation in AI applications for daylily research, including accurate trait prediction and classification.

Broadcast herbicide applications across fallow fields with sparse weed populations can lead to substantial chemical waste, increased costs, and environmental impacts. To address this issue, we developed and evaluated a low-cost, vision-based targeted herbicide spraying system leveraging the Open Weed Locator (OWL) methodology using only the OWL hardware with project-specific custom software. It utilizes simple RGB cameras and image-processing algorithms to detect green vegetation (weeds) against soil backgrounds in real-time, triggering targeted herbicide application only where weeds are detected. The primary objective of this study is to assess the practical effectiveness of this technology in field conditions, focusing on use-cases relevant to fallow weed management scenarios. Field trials are planned to quantify the system’s performance by evaluating weed detection accuracy through metrics such as sensitivity, specificity and precision. Additionally, we will measure the percentage of herbicide chemical savings compared to conventional broadcast methods. Preliminary analyses suggest that this targeted approach significantly reduces chemical input. Detailed evaluations from forthcoming field trials will provide essential data to guide future improvements and support broader adoption of affordable, precision weed management tools.

Accurate plant phenotyping is essential for crop improvement but remains a major challenge, especially when tracking both above- and below-ground traits over time. Traditionally, these traits are measured manually and destructively, limiting data quantity and quality. At the Texas A

Hemp (Cannabis sativa L.) is a remarkably versatile crop with extensive applications in food, fiber, and medicine, offering environmentally sustainable and highly productive raw materials across various industries. Historically cultivated primarily for fiber, modern dual-purpose hemp varieties now present enhanced economic opportunities by enabling the simultaneous harvesting of seeds for grain and stems for fiber. Nitrogen (N) fertilization significantly influences key growth parameters, including plant height, stem diameter, biomass accumulation, and seed yield. However, conventional nitrogen assessment methods are invasive and labor-intensive. To address these challenges, multispectral drone imaging has emerged as a non-destructive alternative, leveraging correlations between nitrogen levels and leaf chlorophyll content to enable rapid monitoring of critical physiological indicators such as assimilation rates, stomatal conductance, and transpiration rates. In 2024, an experiment was conducted at North Carolina A0.70), chlorophyll content and BNDVI (R² = 0.55), stomatal conductance and NDVI (R² = 0.82), and transpiration rate and MCARI (R² = 0.56). In contrast, negative correlations were observed with SIPI2 (R² = 0.69), TGI (R² = 0.39), and additional SIPI2 indices (R² = 0.54 and R² = 0.39, respectively). This study highlights the potential of integrating drone-based remote sensing and machine learning for efficient, non-destructive monitoring in hemp cultivation. By advancing precision agriculture practices, these technologies offer promising pathways to enhance productivity, optimize nitrogen management, and promote sustainability in hemp cultivation.

Nitrogen (N) management is a major challenge in organic vegetable production, aiming to supply sufficient N for optimal yield and quality while minimizing N losses. These challenges are exacerbated by uncertain mineralization patterns of different organic fertilizer products and season-specific impacts on soil N availability. Monitoring plant tissue N dynamics throughout the production season can provide meaningful information regarding fertility management, but sampling plant tissue can be labor-intensive and costly, and lab analysis may be time-consuming. Considering that N from organic nutrient sources is generally not immediately available for crop uptake, the need for reliable tools to rapidly monitor plant N status is paramount in improving N use efficiency, particularly in organic systems. In this study, a hyperspectral imaging approach was explored. Celery samples were collected from two research trials repeated in two production seasons at midseason (approximately 70 days after transplanting; DAT) and final harvests (approximately 110 DAT). One experiment focused on integrated nutrient management practices, comparing celery grown following a sunn hemp cover crop to a weedy fallow (whole plots), and evaluating impacts of composts (subplots) including an unamended control, yard waste compost, vermicompost, and a mixed compost treatment. The other experiment compared ratios of preplant N:in-season N application from 0-100% preplant N (whole plots) under two organic preplant fertilizers (subplots) contrasting in composition and anticipated N mineralization. At each harvest, six plants from each experimental unit were weighed and allocated into representative portions for crop quality analyses on fresh and dry bases. A spectroradiometer with a leaf clip reflectance probe was used to collect leaf reflectance spectra (350-2500 nm) from approximately eight leaves from each experimental unit. Spectra were normalized to a mean of zero and a standard deviation of one across wavelengths. To ensure balanced representation in training and validation, the data were stratified by harvest timing, trial, and year, and randomly split following 80% calibration and 20% external validation distribution. The final model used 19 latent components, explained 76.4% of the variation within external validation data, and had a root mean square error of 0.36. The model can be categorized as providing “approximate quantitative predictions,” and total N content from dried aboveground biomass in the original dataset spans from 1.0 - 4.3 g N/100 g DW. Building robust models using hyperspectral data to predict crop N status under diverse production practices and environmental conditions is an area deserving of continued research in organic vegetable production.

Disaster events such as hurricanes strongly affect pecan crop production due to fallen trees, which generate a significant impact on production areas. An appropriate method to monitor these effects is by counting the fallen trees. This monitoring is commonly performed by a field team. Although accurate, this approach is time-consuming, costly, and non-scalable. Consequently, an image-based approach emerges as a timely opportunity. For instance, aerial images captured by unmanned aerial vehicles (UAVs) present high spatial resolution and coverage, contributing to tree identification. However, relying solely on images remains a laborious and time-consuming task. Consequently, the development of object detection models to automatically identify and count fallen trees emerges as an unprecedented and dynamic method. Therefore, we aimed to integrate UAV image technology and the YOLOv12 detection model for detecting fallen trees in pecan fields. On September 27th, 2024, the hurricane Helene crossed the State of Georgia, USA. On October 7th, 2024, we captured RGB images using a multirotor UAV. We captured high-resolution (~3 cm/px) in eight pecan fields. The images were processed to generate one orthomosaic for each field. Consequently, we counted the fallen trees using assisted image processing for ground truth data. Together, the pecan fields presented more than 200 fallen trees. Subsequently, we trained the YOLOv12 model to automatically identify the fallen trees. Initially, we manually labeled all the fallen trees as “Fallen”. Afterward, we split our dataset into training (70%), validation (15%), and testing (15%). To enhance model performance and generalizability, we implemented data augmentation for the training dataset. As a result, the model achieved identification metrics of fallen trees higher than 70%. Certainly, this approach presents a great opportunity to measure the effects caused by hurricanes, giving farmers the ability to make faster and better decisions regarding their fields. Ultimately, these findings support precision agriculture practices and introduce pecan crops into this context, further strengthening the introduction of technologies into the world of specialty crops.