Progress 08/01/24 to 07/31/25

Outputs
Target Audience:During this reporting period, our primary target audiences included undergraduate students as well as faculty, scientists, and researchers interested in weed management and agricultural techniques that incorporate advanced technologies. We targeted undergraduates to introduce experiential research at an early academic stage through a Course-based Undergraduate Research Experience (CURE) called Introduction to Molecular Modeling. This course integrated programming and machine learning, allowing students to implement a practical computer vision project that effectively differentiated between annual sowthistle (Sonchus oleraceus) and little mallow (Malva parviflora), which are highly problematic agricultural weeds in our region. Accessible computational tools, including the Wolfram Language suite, facilitated broad student engagement and participation regardless of their prior technical expertise. This project also provided the opportunity for four undergraduates to participate in a semester-long independent study course in which they subjected annual sowthistle and little mallow to different abiotic stress factors and created a library consisting of over 2,000 photos to train the initial machine learning model. A hyperspectral imaging component was also added to Co-PI Jani's weed science course, which allowed 45 students to gain direct exposure to hyperspectral imaging cameras and the methods and models behind the technology as it applies to robotic weed control. Scientists and researchers from academia, industry, and USDA-ARS formed another critical audience, primarily engaged through poster and oral presentations at three scientific conferences. These interactions enabled direct dissemination of our research findings, fostered dialogue on computational approaches in agricultural weed management, and established collaborations with experts across disciplines. Faculty, especially those engaged in interdisciplinary research linking soil chemistry, entomology, plant pathology, and machine learning, were also a significant audience. National, regional, and local workshops, seminars, and collaborative meetings allowed us to share methodologies and outcomes, further extending the project's impact. Our research indirectly targets agricultural communities and stakeholders in the California Central Coast region, such as local vegetable and berry farmers, University of California Cooperative Extension agents, and policymakers who benefit from the actionable insights generated. Future direct engagement with these stakeholders is planned to maximize the practical applicability of our findings. Collectively, our targeted efforts enhanced student research capabilities, facilitated academic collaboration, and produced relevant knowledge to support sustainable agricultural practices. Changes/Problems:During this reporting period, there were a few modest but notable changes to the project schedule and scope: Delay in planting a third weed species (purslane): Due to greenhouse space limitations and scheduling constraints, we were unable to plant and image the third weed species originally listed in our goals. This activity has been rescheduled for Fall 2025 and remains a priority for completing the comprehensive weed image dataset. In August, we will travel to a nearby company farm with a known infestation of purslane to dig up young plants and transport them back to campus. The plants will then be potted, subjected to different abiotic stresses, and HSI photos will be taken. Our USDA ARS collaborator is also expected to produce INSV-infected purslane which we will use for HSI analysis as well. Addition of INSV-infected plant imaging: We established a new collaboration with a USDA ARS entomologist who will continue to provide us with virus-infected weed specimens. This represents a significant enhancement to the original experimental design and will allow us to expand our data collection to include spectral signatures associated with plant disease. Machine learning efforts extended to new data types: While the original project focused on RGB and HSI data from healthy weeds, we are now incorporating MATLAB-based analysis of infected HSI images. This shift expands the computational demands and technical scope of the project but is fully aligned with the project's goals and long-term impact. There were no changes to the approved Data Management Plan, and no deviations from protocols related to biosafety, human subjects, or animal use. No issues have impacted the rate of expenditure or required additional reporting under the award's Terms and Conditions. What opportunities for training and professional development has the project provided?This project provided extensive training and professional development opportunities for undergraduate students, particularly those in their first year of STEM coursework. A total of 14 students participated in a Course-based Undergraduate Research Experience (CURE), where they engaged in a research project focused on computer vision and machine learning tools to distinguish between invasive weed species linked to significant crop losses in Monterey County. Students received hands-on training in the following areas: Hyperspectral Imaging (HSI): Students learned to operate and troubleshoot HSI cameras, which capture detailed spectral information across a wide range of wavelengths. These systems provide powerful insights into plant physiology and disease markers not visible in standard RGB imagery. Machine learning and image classification: Using convolutional neural networks and tools such as the Wolfram Language, students were trained to develop and evaluate models capable of distinguishing weeds under varied growing conditions. Data management and annotation: Students organized and labeled large image datasets for machine learning pipelines, learning best practices in dataset structure, metadata handling, and version control. Research ethics and integrity: Training was provided on responsible conduct in research, including data handling, authorship, reproducibility, and ethical use of AI. Scientific communication: Students practiced writing abstracts, developing research posters, and delivering oral presentations. Several presented at campus and regional conferences. In addition to the course-based work, two students were hired full-time over the summer to expand the research by integrating and analyzing hyperspectral imaging data using MATLAB. MATLAB is a professional scientific computing environment widely used in academia and industry for data analysis, image processing, and algorithm development. Its built-in support for matrix manipulation, spectral analysis, and machine learning made it an ideal platform for developing scripts to process and interpret complex HSI data. Students received targeted mentorship on MATLAB coding practices, image preprocessing workflows, and the development of custom analysis pipelines for high-dimensional spectral datasets. As mentioned previously, four independent study students and 45 weed science students gained exposure to HSI applications to weed management through their participation in this project. The independent study student has no prior research experience. Along with some of the students listed above, they gained exposure to HSI cameras and related software by attending on campus intensive trainings provided by an engineer from the camera's manufacturer. Having collected over 2,000 images of RGB photos, they also were tasked with learning how to efficiently organize large data sets, all while managing the weeds growing under different stress conditions. A number of the 45 weed science students who learned about HSI application to weed management completed summer internships with companies currently using or interested in adopting robotic weeding. As mentioned earlier, this reporting period also marked the beginning of a new collaboration with a USDA ARS entomologist, who has begun providing INSVI-infected weed samples for future analysis. This partnership enhances the professional network available to students and creates a bridge between academic research and federal agricultural priorities. Altogether, this project has served as a gateway for students to gain high-impact, career-relevant research experience in agricultural technology, image analysis, and ethical AI. Many of the skills developed--such as coding, data annotation, and scientific reasoning--are widely applicable across STEM fields and contribute to a strong foundation for future research opportunities, graduate study, or industry careers. How have the results been disseminated to communities of interest?The results of this project have been actively disseminated through student-led conference presentations, designed to engage both scientific and broader agricultural communities. Undergraduate researchers served as the primary communicators of this work, helping bridge the gap between academic research and real-world agricultural challenges. During this reporting period, students presented posters and gave talks at major scientific and professional meetings, including: American Chemical Society (ACS) National Meeting, San Diego, CA, Spring 2025 "Protecting Crop Health: Machine Learning for Weed Detection with Hyperspectral Imaging" (Supreet Gandhok*, Trevor Pollock*, Eduardo Lopez*, Owen McMillan*) This national platform allowed students to share their work with a diverse audience of chemists, educators, and industry professionals, highlighting applications of machine learning and HSI in precision agriculture. California Plant and Soil Conference, Visalia, CA, Spring 2025 "Using Machine Learning to Identify Problematic Central Coast Weeds Experiencing Environmental Stress" (Eduardo Lopez*, Trevor Pollock*, Supreet Gandhok*) This outreach was targeted specifically at California's agricultural community, includinggrowers, extension agents, industry professionals, crop consultants, professors, USDA ARS and NRCS personnel, and policy makers. The presentation focused on regionally relevant weed species and drew attention to the potential of AI to assist in local crop protection strategies. Western Society of Crop Science Annual Meeting, Walla Walla, WA, Summer 2025 Hyperspectral Imaging of Annual Sowthistle Experiencing Abiotic Stress (Eduardo Lopez)* This talk focused on early data integration between RGB and HSI modalities and its potential for enhancing early weed and disease detection tools. This meeting is regional in scope and brings together stakeholders from throughout the western USA, including growers, extension agents, industry professionals, crop consultants, professors, USDA ARS and NRCS personnel, and policy makers. Presenting the project at this meeting expanded its scope to include stakeholders outside of California. In addition to formal conferences, the project was featured at campus events such as undergraduate research symposia and classroom presentations, helping to raise awareness among students not already involved in research. These activities directly support public understanding of science and encourage broader student interest in agricultural technology, artificial intelligence, and environmental sustainability. Importantly, by having undergraduates serve as the public face of this research, we are cultivating the next generation of agricultural scientists and data-driven problem solvers. Their visibility and success in public forums help demystify STEM careers for peers and community members, particularly those from groups historically underrepresented in science. Future dissemination efforts will focus on more in-depth engagement with grower networks and extension professionals as the applied outputs of this work mature. What do you plan to do during the next reporting period to accomplish the goals?During the next reporting period, we will undertake the following actions aligned with the three core project goals: Goal 1: Development and Validation of Machine Learning Models We will significantly expand our efforts in building and validating machine learning models using Hyperspectral Imaging (HSI) data. Specifically, we will begin developing classification models to distinguish between INSV-infected and healthy weed specimens, made possible through the previously mentioned collaboration with the USDA ARS entomologist, who will continue providing access to infected plant material. We will increase our use of MATLAB, a scientific computing platform well-suited for HSI data analysis. MATLAB supports advanced spectral processing, dimensionality reduction, and model training workflows, enabling us to explore both traditional classification and deep learning approaches. We will begin comparing model performance between RGB and HSI modalities, focusing on identifying spectral features that may allow early detection of weed-borne disease. Goal 2: Comprehensive Data Collection and Analysis We will continue collecting HSI data for non-infected weed species, including annual sowthistle and little mallow, under multiple abiotic stress conditions. We will introduce and image a third weed species, purslane (Portulaca oleracea) identified in our original goals, which was not planted during the current period due to timing constraints. This species will be grown and imaged during Fall 2025, with both RGB and HSI data collected across its early growth stages. We will expand our spectral dataset to support future model training and to begin identifying spectral signatures associated with INSV infection and abiotic stress. This step is critical for establishing the spectral distinctiveness of problematic weeds and for building tools that support real-time detection. Goal 3: Student Training and Collaborative Engagement We will continue to engage undergraduates through CURE courses and independent research, with a focus on students in Agricultural Plant and Soil Science and related programs. Two students will be supported during summer and academic year research to specialize in HSI data analysis, MATLAB scripting, and image processing workflows. Training will also include data ethics, AI literacy, and communication of scientific results to both technical and non-technical audiences. We will strengthen our collaboration with USDA ARS by incorporating infected samples and co-developing imaging protocols that reflect real-world agricultural conditions. In addition, we plan to submit one or more peer-reviewed manuscripts reporting the results of our model development and dataset creation. These efforts will ensure that our findings reach academic, technical, and applied agricultural communities, fulfilling the broader impacts of the project.

Impacts
What was accomplished under these goals? This project addresses a major challenge facing specialty crop production (e.g., lettuce, broccoli, artichokes, etc.) in the California Central Coast region: environmental and disease-related stress factors. Limited water availability, especially in combination with heat stress, can cause tremendous damage to crops and is uncommon in the region in part due to dwindling water supplies and periodic droughts. Although weeds are often ignored under such scenarios, they can also exhibit these environmental stresses. It is important that AI models used by robotic weeders are exposed to weed species, such as annual sowthistle and little mallow, under these different stress scenarios. This is so that weeds growing under these conditions can be identified using AI models and weeds can then be terminated. In Monterey County alone, annual sowthistle and little mallow are not only difficult to control, but they also serve as hosts for Impatiens Necrotic Spot Virus (INSV), which cost farmers more than $150 million in lost lettuce crops in a single year. Traditional weed control methods are expensive, labor-intensive, and often harmful to the environment due to dependency on synthetic herbicides. These same weed species and INSV are also widespread in Yuma, Arizona, which is the leader in winter lettuce production in the USA. Thus, interest in this project extends beyond the Central Coast of California. Our project aims to reduce this burden by developing new, precise tools to identify and manage weeds experiencing environmental and disease stress that may be encountered in the field using artificial intelligence (AI) and computer vision. During this reporting period, we accomplished the following: 1. Developed the first curated image dataset for two high-impact weed species in California agriculture We photographed over 800 weed specimens (annual sowthistle and little mallow), each grown under a range of realistic agricultural conditions, including drought, overwatering, high/low fertilizer, and physical damage. Each image was enhanced using computer algorithms that mimic real-world variability like changes in light, blurring, and partial visibility due to blocking by crops. This resulted in a final dataset of over 23,000 images used for model training and testing. 2. Trained and evaluated three advanced machine learning models (ResNet-50, ResNet-101, DenseNet-121) These models were trained to identify and differentiate between the two weed species across 10 visually similar categories. The best-performing model achieved a classification accuracy of over 96% under variable lighting and environmental conditions. DenseNet-121 and ResNet-101 consistently outperformed simpler models, with ResNet-101 providing the most reliable predictions. 3. Integrated this research into a classroom experience We implemented this work into a Course-based Undergraduate Research Experience (CURE) for first-year STEM students. Students learned programming and machine learning to build their own image classification tools. By the end of the course, they could train models to correctly identify weed species--a skill that translates directly to real-world agtech applications and employment opportunities in the region. 4. Shared results with the scientific community The team presented posters at three regional and national research conferences, reaching scientists, educators, and potential collaborators. Team leaders collaborated with local USDA-ARS scientists to take HSI photos of several weed species, including annual sowthistle and little mallow, that were inoculated with INSV to expand the library dataset. A manuscript based on this work and coauthored by some of the project's undergraduate researchers is currently under peer review, and data will be made available to other researchers upon request. Outcomes and Impacts Improved knowledge: Students gained hands-on experience in AI and agricultural technology, cultivating weed species under imposed stress factors, preparing them for careers in science, technology, and agriculture. New tools for agriculture: We developed a reliable, localized weed identification system that can be incorporated into existing automated farming tools like drones and robotic weeders. Environmental and economic benefits: Early and accurate weed detection can reduce pesticide use in two ways: 1. Terminating young weeds before they establish and require larger amounts of herbicides. 2. Reducing incidence of INSV in the field and thereby limiting pesticides needed on infected lettuce that is also susceptible to secondary infection by other pathogens in its weakened state. Early detection can also reduce labor costs and crop losses--ultimately benefiting farmers and reducing environmental harm. Research infrastructure: The dataset and models created during this period are now available for future development of disease management systems and agricultural robotics. This work matters to farmers, agtech developers, students, and researchers. By building practical tools and training the next generation of problem solvers, we are contributing to California workforce development while making agriculture more efficient, sustainable, and resilient.

Publications