Duration: 10/01/2026 to 09/30/2031

Administrative Advisor(s):

NIFA Reps:

Non-Technical Summary

Statement of Issues and Justification

The need as indicated by stakeholders

The competitive nature of modern agriculture demands agribusiness firms innovate and adapt quickly to capture benefits from advances in new technology such as artificial intelligence (AI). Throughout the value chain, there is a critical need to increase efficiency and protect the bottom line. Although currently lagging behind other industries, agriculture is forecasted to experience a “digital revolution” over the next decade.  The market value of AI in agriculture is expected to grow at a compound annual growth rate of 22% per year, reaching $8 billion by 2030, according to a recent report by Insight Partners. Growers, for example, are constantly exploring the best opportunities to increase yield and profit as expressed by Iowa corn and soybean producers at the recent NSF Convergence Accelerator Workshop for Digital and Precision Agriculture. Participants strongly voiced the need for new technology to answer “what is my best opportunity?” to enhance crop productivity and strengthen their bottom line.

Based on the National Science & Technology Council (2019) report, the American Artificial Intelligence Initiative was established in 2019 to maintain American leadership in AI and ensure AI benefits to the American people. The initiative sets up long-term investments in AI research as one of the priorities. Following the initiative, USDA/NIFA invests in many AI-related programs such as the Data Science for Food and Agricultural Systems (DSFAS) and other programs for crop and soil monitoring systems, autonomous robots, computer vision algorithms, and intelligent decision support systems.

AI allows computers and machines to carry out tasks without human cognition. AI includes machine learning (ML) and deep learning (DL). Machine learning is a data analysis method to imitate human learning. Examples of using ML in agriculture include crop production (e.g., yield prediction, pest and disease detection, harvesting, and phenotyping), livestock farming (e.g., behavior and health monitoring), postharvest processing (e.g., quality inspection, grading and sorting), and agricultural robotics (e.g., navigation, self-leaning, human-machine synergy, and optimization). Deep learning is a subset of machine learning and consists of deep neural networks to mimic human brain functions. Recently, deep learning has driven huge improvements in various computer vision problems, such as object detection, motion tracking, action recognition, pose estimation, and semantic segmentation (Voulodimos et al., 2018). Internet of Things (IoT) is a technology that enables data acquisition and exchange using sensors and devices through a network connection. IoT tends to generate big data, which requires AI to make valuable inferences. 

Crop growth is a complex and risky process and difficult for producers to analyze in isolation. The emergence of information technology has developed a large amount of data, i.e., big data, that can be analyzed utilizing AI to provide valuable decision support to producers, particularly large-scale operations. AI can provide predictive analytics for growers to better manage risks through improved preparation and response for unexpected events such as severe flooding and drought. For instance, the Colorado-based company, aWhere, uses machine learning algorithms in connection with satellites and 1.9 million weather stations (virtual) to predict daily weather patterns for clientele, including farmers, crop consultants, and researchers. Improved crop and irrigation planning assists a global network of growers in reducing water usage with a particular focus on the impacts of climate environmental changes.

Robotics and visual machine learning platforms enable automation of critical labor activities such as field scouting, weeding, and harvesting. Pest control companies have begun using aerial drone technology to cut costs in the labor-intensive practice of scouting pests and diseases. According to Brian Lunsford of the Georgia-based Inspect-All company, “In 2016, we performed our first paid drone inspection after months of testing. Most importantly, we wanted to make sure our flight operators could safely fly our drones and provide our customers with substantial value, while at the same time being mindful of privacy concerns.”  As reported by Protein Industries Canada, AI-assisted pest and disease monitoring systems can save pesticide use up to 95% and reduce costs by $52 per acre. Ground robotic weeding systems powered by AI and big datasets have started to be commercially adopted for precision farming, which can recognize weeds and kill them site specifically, substantially reducing the need for both manual labor and herbicides.

Emerging AI technologies are expected to reach a broader spectrum of the value chain. While prior technologies focused primarily on field crop commodities, AI algorithms will enable specialty crop industry stakeholders to optimize production, improve harvest efficiency and cut labor costs by using AI-enabled automation or robotics systems. Early adopters such as John White of Marom Orchards support the use of AI in response to the need “to pay wages, organize visas, housing, food, healthcare and transportation” for a large number of workers and to address critical labor shortages due to “hard, seasonal work and other crops can pay higher wages. Young people all over the world are abandoning agricultural work in favor of higher paying, full-time urban jobs.” Small, labor-intensive farm operations can thus expand production opportunities by reducing harvest losses by 10% while reallocating freed-up labor to alternative enterprises, alleviating concerns over expected labor shortages that are expected to reach $5 million by 2050, according to Marom Orchards.

Downstream on the value chain, AI is also expected to have a strong demand over the coming decade. A recent article in Food Online lists three new types of AI technology to improve supply chain management, including: (1) food safety monitoring and testing of product along the supply chain; (2) improved marketing analysis of price and inventory; and (3) a comprehensive “farm to fork” tracking of product. In the food processing industry, AI algorithms are being used at modern packing facilities for food inspection, grading, and sorting. For instance, the Spectrim (an optical sorter) from TOMRA Food has leveraged deep learning for enhanced grading precision of fruits. AI reduces labor time compared to manual sorting, enhances quality and consistence of products adelivered to consumers, and reduces postharvest losses. To fine-tune product development and optimally satisfy consumer preferences, startup companies such as Gastrograph AI use machine learning to assist their clientele in fine-tuning product development. AI is expected to be in high demand to improve hygiene in both manufacturing plants and restaurants. The use of AI in the cleaning of manufacturing equipment is projected by the University of Nottingham researchers to reduce cleaning costs by up to 40%.

 

The importance of the work, and what the consequences are if it is not done

There are pressing challenges within the country’s agroeconomic systems. These challenges directly impact growers, rural communities and consumers, and include:

• Growing enough agrifood to meet growing population demands
• Rising costs
• Invasive pests and weeds
• Plant and animal diseases
• Excessive livestock mortality
• Land and water supply degradation
• Changing environmental conditions that harm production
• Shortage of labor
• Reduce waste and ensure safe food processing
• How to increase production while reducing land use

AI has potential to address and significantly ameliorate many of these agroeconomic challenges. 

• Improve efficiencies and reduce pollution through targeted interventions
• AI tools can analyze large, diverse datasets (remote sensing, IoT, genomics, weather, management) to guide better decision-making, increase yield, and reduce input costs.
• Precision AI technologies can reduce water, fertilizer, and pesticide use, cutting waste and environmental pollution.
• Predictive analytics can detect crop stress, pests, and diseases earlier, improving resilience to shocks such as droughts, floods, or emerging pathogens.
• Robotics and computer vision can support harvesting, scouting, grading, and sorting, which are critical in the face of farm labor shortages.
• AI tools can non-invasively monitor livestock behavior and health indicators, improving welfare while reducing unnecessary treatments and antibiotic use.
• The project will train a new generation of scientists, extension agents, and producers in AI-driven agriculture, ensuring adoption and long-term capacity.
• AI tools can nondestructively detect food fraud, defects in produce, classify food, detect and quantify contaminants such as allergens in food.

 The AI work proposed here will:

• Deliver field-ready solutions that improve profitability, resilience, and sustainability across U.S. agriculture.
• Enable farmers to manage risks and adapt to extreme weather events more effectively.
• Build a strong foundation for U.S. leadership in agricultural AI, supporting competitiveness across global markets.
• Advance environmental stewardship by reducing agriculture’s footprint on soil, water, and the atmosphere.

 If this work is not done:

• There will be a significant opportunity cost.
• Farmers do not have the resources to perform or fund this work.
• Other countries that develop AI solutions for their farmers will gain a competitive advantage.
• Inefficient use of resources will persist, continuing to drive environmental degradation and loss of applied nutrients to water and air.
• Labor shortages and farm-level vulnerabilities will deepen, threatening rural community viability and national food security.
• Shortage of food, less safe foods, more wastes are produced from agrifood production processes, and supply chain challenges.

The importance of this project lies in that it will help tackle multiple pressing challenges that we are currently facing in the agroecosystem. The current AI technologies are not explicitly tuned for agroecosystem, which causes problems such as low accuracy in prediction, inefficiency in using computer resources, inefficiency in data management, and not being cost-effective for most agricultural crop productions. Also, the lack of next-generation farmers and workers in this area will be a major bottleneck for adopting and applying AI technologies. Included in this project, multiple AI-centered projects will be conducted at multiple states in the southeast U.S. to develop AI tools suitable for specific applications important to the improvement of production and sustainability of the agroecosystem. The project will also assess the feasibility of different AI technologies and showcase the value of those technologies to stakeholders to improve AI adoption. Additionally, this work will help develop the workforce for the future agricultural production system. Tasks in this project should be completed quickly and efficiently to ensure that production in agriculture meets the global needs in the near future and ensures the sustainability of the agroecosystem. It is also essential that technologies in agriculture must keep up with technologies in other fields to attract more talented people to ensure workforce sustainability. 

   

The technical feasibility of the research

The proposed research is technically feasible, benefiting from recent advances in AI, remote sensing, high throughput sensing technologies, digital twins, edge computing, and integrated data platforms. The team has access to robust computational resources, including cloud-based platforms and local high-performance computing clusters, that are capable of handling large-scale, heterogeneous datasets typical in agricultural systems. These resources support efficient model training, geospatial analysis, deployment of decision support tools, digital twin simulations, and near-real-time decision-making.

Machine learning algorithms, including ensemble methods, deep neural networks, and explainable AI techniques, are now mature enough to be applied across multiple data streams (e.g., optical remote sensing imagery, IoT sensor data, drone/satellite spatial data, and climate models) to provide real-time information and sampling efficiency. Open-source deep learning libraries such as TensorFlow, PyTorch, XGBoost, and Earth Engine facilitate scalable development and deployment of predictive models. Additionally, software for precision agriculture (Agisoft, ArcGIS Pro QGIS, OpenDroneMap) and remote sensors allow for real-time integration of spatial and temporal data for efficient, local and large-scale crop monitoring.

The research team is composed of multidisciplinary experts in agronomy, soil science, entomology, plant pathology, statistics, and computer science, and is already working with regional and national stakeholders. Existing data pipelines from previous and ongoing projects can be readily extended, reducing time-to-deployment for new analytical models. Furthermore, the availability of low-cost IoT sensors, drones, and open-access satellite data ensures that experimental designs can be implemented at both plot and landscape scales.

Taken together, the combination of mature technologies, available infrastructure, and interdisciplinary expertise ensures the technical feasibility of the project.

 

The advantages  for doing the work as a multistate effort

Each state has different experts in different research areas of AI and applications. Working as a multistate team creates more opportunities for collaborating in various disciplines from other states. From the survey completed in May 2021 at SAASED Institutions (SAASED, 2021), only 46% of the respondents have developed a partnership with other institutions. The survey also found that it is less likely that researchers know what other institutions are doing, and there is not much coordination among them. Therefore, this multistate project will facilitate more productive collaboration through organized coordination to complete the tasks proposed in this project.  

At the initial proposal development stage in May-July 2021, we had meetings every one or two weeks, and there were 15-20 participants for most of the meetings. As a group, we came up with the title, defined objectives, and established the writing team for each specific objective. For each objective, a leader was chosen to lead the writing activities. All members were committed to accomplishing their assigned tasks of writing an introduction, literature review, and detailed activities, along with outputs, outcomes, and milestones, which resulted in this proposal.

Since the creation, we have seen a steady growth of this multistate group with new members participating each year. This group has been meeting twice a year, one at the AI in Agriculture Conference created and sponsored by this group, and the other meeting specifically for this group. At these meetings, participants from different institutions discuss their research activities, major findings, current issues, funding opportunities, potential collaborations, and future directions. A listserv of emails was created to facilitate efficient communication among the participants. An online cloud folder was created to share and maintain data and information. The group has fostered extensive collaborations across institutions and impactful achievements in advancing AI in Agroecosystems.  

 

What the likely   will be from successfully completing the work 

This multistate project could significantly improve farming practices and productivity. With AI tools, growers would be able to identify the best opportunities to increase yield and profitability, using predictive analytics and real-time data. Technologies such as machine learning and deep learning would help detect crop diseases, stress, and pests earlier, allowing for faster and more targeted responses. For large-scale operations, AI would simplify the complexity of managing crops by turning big data into actionable insights. 

Automation would also play a major role in addressing grand labor challenges. Robotics and visual machine learning platforms could take over labor-intensive tasks such as scouting and harvesting, which are especially important given the growing shortage of agricultural workers. Small/medium-sized farms could also benefit by reducing harvest losses and reallocating labor to other areas. They will be able to expand production without increasing workforce demands.

Beyond farming, AI would strengthen the entire agricultural value chain. It could improve postharvest handling, food quality and safety, reduce waste, and support product development based on consumer preferences.

Environmental sustainability would improve as well. AI-driven precision agriculture could reduce the use of water, fertilizers, and pesticides, helping farms lower their environmental impact while maintaining productivity. These technologies would support better planning for extreme weather events such as droughts and floods, making farms more resilient to climate change.

On a national level, the project would support goals set by America’s AI Action Plan, published in July 2025, including (1) Build World-Class Scientific Datasets, (2) Build an AI Evaluations Ecosystem, and (3) Enable AI Adoption. This project would contribute to achieving these goals by developing AI tailored to agriculture, helping the U.S. maintain leadership in this space. By addressing current limitations in cost, accuracy, and data management, the project would help make AI tools more accessible and effective for producers.

Finally, the project would help build a skilled workforce. By integrating AI and robotics into education and extension programs, it would prepare the next generation of agricultural professionals. Multistate collaboration would also foster innovation across institutions, leading to joint publications, shared datasets, and coordinated outreach efforts.

 

Related, Current and Previous Work

Artificial intelligence (AI), including machine learning (ML) and deep learning (DL), has emerged as a powerful approach for handling large, heterogeneous sensor datasets in agriculture. Among DL approaches, convolutional neural networks (CNNs), deep neural networks (DNNs) (e.g., long short-term memory, LSTM), vision transformers, vision foundational models, and recurrent neural networks (RNNs) have received great attention in agriculture.

 

AI for Crop, Animal, and Food Production

Sustainable crop production faces persistent challenges in managing biotic stressors, such as weeds, insects, and diseases, as well as abiotic stresses arising from variable environmental conditions. Sensing technologies—particularly imaging-based systems—offer promising tools for monitoring crop health, pest pressure, and field variability. AI-based image analysis and modeling have demonstrated strong performance in key vision tasks, including crop stress detection, disease identification, and pest monitoring, enabling more precise and timely management interventions.

Accurate in-season yield prediction is essential for optimizing input management, harvest planning, crop insurance, and marketing decisions. Crop yield variability results from complex interactions among management practices, climate, water availability, soil properties, genetics, pests, diseases, and weed pressure, making reliable prediction under real-world field conditions inherently challenging. AI-based yield prediction has been successfully demonstrated across major agronomic crops, including wheat, sorghum, soybean, corn, and rice. Specialty crops present additional complexities due to variable geometry, asynchronous maturity, and multiple harvest cycles. Nevertheless, AI-driven image analysis and neural networks have enabled early yield prediction for fruits and vegetables such as apples, blueberries, strawberries, peppers, tomatoes, apricots, and eggplant, offering substantial improvements in accuracy and efficiency over traditional methods.

Advances in ML have also enabled individual-level animal monitoring for health and welfare assessment. Unsupervised techniques are commonly used for animal segmentation, while supervised learning models support posture, gait, and behavior recognition. Applications include lameness detection in dairy cattle, thermal comfort classification in pigs, aggression monitoring, and estrus identification. However, most existing studies focus on large animals in confined environments; research on poultry remains limited and is largely conducted under laboratory conditions, although promising results have been reported for broiler health assessment, gait scoring, and behavior monitoring using computer vision and ML.

In addition, AI-driven sensing technologies provide nondestructive alternatives to conventional postharvest food quality evaluation methods. These approaches enable rapid assessment of physicochemical properties, defects, contamination, and pest infestation in agricultural products. When integrated with blockchain technologies and predictive analytics, AI-based systems have the potential to enhance transparency, efficiency, and waste reduction across food supply chains.

 

AI for Improved Robotic Systems

AI plays a central role in agricultural robotics, particularly in scene interpretation, object detection, navigation, vision-based control, and fleet management. Deep learning has significantly improved fruit detection and localization, which are foundational for robotic harvesting and automated yield estimation. Architectures, such as You Only Look Once (YOLO) series, real-time detection transformers, and the Segment Anything Model (SAM), have been widely adopted, though data annotation requirements remain a bottleneck.

Navigation in orchards poses additional challenges compared to open-field crops. Machine vision, light detection and ranging (LiDAR), GPS-based sensor fusion, and multi-sensor approaches have all demonstrated success in row-following and path planning. Vision-based control systems—particularly closed-loop visual servoing—enable precise fruit manipulation, though controller stability and robustness remain active research areas. Beyond individual robots, AI also supports fleet-level optimization, diagnostics, and task planning, offering pathways to improve system reliability and economic viability.

 

AI for Natural Resources Scouting and Monitoring

Machine learning has accelerated the analysis of soil and environmental data, supporting applications in soil carbon mapping, soil health assessment, and nutrient modeling. Deep learning has also shown strong performance in water quality monitoring and harmful algal bloom detection using remote sensing data, particularly when multimodal and spatiotemporal datasets are integrated through CNN–LSTM architectures.

 

AI for Plant Phenotyping and Genotyping

Deep CNNs have revolutionized image-based plant phenotyping, enabling classification, regression, segmentation, and object detection tasks that were previously difficult using traditional methods. Applications span species identification, disease detection, organ counting, biomass estimation, lodging assessment, and 3D reconstruction using color, spectral, LiDAR, and other sensing modalities. Despite progress, the limited availability of open-source datasets remains a key barrier to rapid algorithm development and benchmarking.

 

Cross-Cutting Needs: Standardization, Economics, and Education

While many AI technologies demonstrate strong technical performance, fewer studies address their economic, environmental, and social impacts. Economic surplus models, risk analysis, and biophysical simulation tools provide frameworks for evaluating adoption outcomes but are underutilized in AI-focused research. Additionally, the “black-box” nature of AI remains a barrier to adoption, underscoring the need for education, explainability, standardized testbeds, and interdisciplinary training.

Objectives

1. Develop AI-based approaches for agroecosystems production, processing, & monitoring
   Comments: Sub-objectives: a. AI tools for food, crop, and animal production and processing; b. AI tools for autonomous system perception, localization, manipulation, and planning for agroecosystems; c. Natural resources scouting and monitoring; d. Phenotyping and genotyping;
2. Data curation, management, accessibility, security, and ethics
   Comments: Sub-objectives: a. Create open source agricultural datasets following FAIR (Findable, Accessible, Interoperable, Reusable) principles; b. Data Standardization and testbed development; c. Agricultural data governance, ethics, and privacy
3. AI adoption (technology transfer) and workforce development
   

Methods

Objective 1: Develop AI-based approaches for agroecosystems production, processing, & monitoring

Obj. 1a. AI tools for food, crop, and animal production and processing

A. Introduction

The advancements in AI tools and the improved computational power make it easier to handle and derive meaning from large amounts of agricultural data. The ability to process large amounts of data in real-time or near-real-time would help improve the productivity of crops and livestock with precise management.

AI tools can be used for various aspects of crop production, including crop status/growth monitoring, nutrient and water stress detection, insect pest/disease identification and management, weed detection and management, precision spraying, and postharvest processing and quality evaluation. In livestock production, AI tools can also be used for welfare assessment, behavior analysis, status monitoring, and assisting with management instructions.

In both agricultural and livestock production, environment and facility settings often vary from one and another. For instance, crop yield depends on several factors that vary from field to field. A yield prediction model developed for a particular crop on a specific field will likely not work in another field. As the amount of data collected from satellites, aerial and ground sensors, weather forecasting, and plant genomics increase, new and more accurate, reliable, and robust yield prediction and quality evaluation tools are needed. In addition, livestock systems can have vastly different housing facilities with different floor layouts, pen designs, feeders, drinkers, herd size, stock density, and so on. A behaviour recognition model trained using data from a particular facility may show declined performance when applied in a different setup. Livestock may express different levels of heat/cold stress depending on the geolocations, stock densities, and housing design. This can result in an AI model developed to aid with decision making to have inconsistent performance when applied to different sites.  It is critical to develop AI tools with enhanced generalization ability to account for the variations. This may be achieved by incorporating a larger amount of data and including a wider spectrum of data when developing the AI tools. For example, it can be critical and beneficial for crop yield prediction models to incorporate additional data such as historical yield, images from satellites and drones, surveys of soil fertility and electrical conductivity and topography data, and historical and current weather that are likely accessible by farmers. Similar principles can be applied to livestock farming. For instance, adding vulva swollenness features on top of behavior features showed potential to enhance AI estrus detection performance regardless of different herd size and stock density.

B. Detailed activities/procedures

The team will conduct research in the following areas: crop and animal status/growth monitoring; yield estimation/prediction; biotic or abiotic stress detection and management; and agrifood quality evaluation. Various imaging and non-imaging sensors will be used to collect data on crop, food, and animal status and growth. Data is expected to be commonly available to farmers will be collected on a cotton field, including soil electrical conductivity, soil type from a soil survey, historical yield data, satellite data, UAV data, historical and predicted weather data, etc. AI-based models will be developed to use these data in an attempt to predict yield with high spatial precision.

Color images will be acquired to monitor various crop growth stages and to detect and differentiate crop nutrients, yield, diseases, plant pathogens, and weeds. Thermal imaging will be employed to identify stressors in the field. The integration of these imaging modalities will enable early diagnosis of stressors and facilitate timely, targeted management strategies to reduce crop losses. While hyperspectral imaging is too costly for widespread field deployment, it will be utilized to identify key wavelengths for specific stressors (e.g., plant pathogens) in the greenhouse and for the development of more affordable multispectral sensors. Additionally, project efforts will focus on transforming high-resolution RGB data into hyperspectral-like multispectral data, with applications in cost-effective handheld devices for qualitative assessment of agrifood systems and disease phenotyping (Oloyede and Adedeji, 2025). As an example, AI models such as YOLO and VGG are being evaluated for their potential to accurately rate root rot severity caused by Rhizoctonia solani in soybean, a prevalent pathogen across all soybean-producing states in the U.S. These models, once validated, may assist in high-throughput disease phenotyping by enabling rapid, objective, and consistent assessment of disease symptoms at scale.

Diseased plants often show chlorophyll degradation—ya breakdown in green color—which alters reflectance in the RGB, near-infrared (NIR), and shortwave infrared (SWIR) spectral bands (Krezhova et al., 2017; García-Vera et al., 2024). AI systems using machine learning techniques—such as convolutional neural networks (CNNs), support vector machines (SVMs), k-nearest neighbor (KNN), and partial least squares regression (PLSRs)—can accurately classify plants as healthy and diseased (García-Vera et al., 2024). These systems enable real-time disease monitoring through drones, sensors, and smartphones, reducing the need for manual scouting and allowing for timely, targeted interventions. AI also enhances predictive modeling by analyzing environmental and agronomic data to forecast disease outbreaks, support decision-making, and optimize pesticide use. For instance, Zanin et al. (2022) demonstrated that real-time, site-specific pesticide application using precision sprayers and AI-enabled UAV surveillance significantly reduced pesticide cutting costs by 2.3 times compared to whole-field spraying. Together, these advancements position AI as a vital tool for advancing precision agriculture and ensuring more resilient, efficient, and sustainable management systems.

While foundational AI research is not the intent of this project, these types of data will be used to apply AI algorithms for detecting and monitoring crop status. Various AI algorithms will be used, including convolutional neural networks (CNN), region-based CNNs, ResNet, You Only Look Once (YOLO), single shot multibox detector (SSD), as well as newer AI architectures like Vision Transformers (ViTs), SAM, and diffusion models have gained traction.

A major difficulty in processing agricultural data is the fact that the data used are collected at specific, different times and locations. For example, an instance of remote-sensing data may be taken during the flowering period in the current season, while a set of historical yield data may have been taken at the end of the season two years ago. Data collected at different times can be represented as a sequence, sequence learning methods can be applied to involve (1) a network into sequences of arbitrary length that can be fed one element of the sequence per time step, and (2) a network that can remember important events that happened many time steps in the past. A variety of different recurrent neural networks will be considered.

Another concern with outdoor image data is that the data is heavily affected by varying illumination. We will train AI algorithms with images of varying illumination to explore whether AI algorithms can handle them. If not, AI algorithms will be trained with different illumination conditions and used separately for individual illumination conditions.

In food processing applications, quality evaluation, grading, and sorting of agricultural products are critical tasks at packing facilities to ensure the separation of high-quality, marketable grades of products and remove culls (low-quality, defective products) that are inferior or unmarkable. Furthermore, in the age of precision agriculture and traceability, commodity quality can conceivably be traced back to field position to enhance spatially variably crop management. Human inspection is routinely used for grading and sorting food products (e.g., sweet potato), especially when sorting for defects. This process is the most intensive labor during postharvest handling. In addition, the human inspection may suffer from inconsistency and variability in selection induced by human subjectivity and other physiological factors. Moreover, packinghouses face a significant challenge of retaining well-trained workers. Hence there is a pressing need to develop automated quality evaluation, grading, and sorting systems. AI-based machine vision will be investigated as a means to automate postharvest grading and sorting on production lines (Blasco et al., 2017) to offer improved efficiency, objectivity, and accuracy.

In addition to post-harvest quality evaluation, non-destructive testing of foods is needed for pest detection in produce, adulterant detection, etc. Near-infrared sensing and other sensing technologies like acoustic sensors are being developed to improve AI deployment for solving problems related to food quality assessment and safety assurance (Adedeji et al., 2024; Rady and Adedeji, 2020; ,Ekramirad et al. 2024). Activities in these areas will include developing multispectral models that are based on AI tools like machine learning and deep learning from near-infrared (HSI, CV, etc.), acoustic, and other sensors.

The team will also conduct research in animal health and welfare assessment using AI-based machine vision . Detailed activities may include detection of broiler behaviors associated with key welfare indicators, estimation of cattle and broiler body weight, and respiratory rate evaluation of swine.

Behavior patterns can reflect the health and welfare of an animal. Numerous of studies demonstrated ways to reliably conduct behavior recognition/classification. One of the remaining challenges is the complex background environment that can vary from farm to farm. To address this issue, we will explore different automated background removal algorithms, such as Segmentation Anything Model and frame subtraction, to retain pixel information related to the animal and discard all unrelated pixel information. We will examine whether, after filtering background information, a behavior recognition model can be applied to different farming facilities while maintaining satisfactory performance.

Bodyweight is an important factor associated with many management practices in livestock production and a good indicator of animal health (Uluta & Saat, 2001). Previous studies used cattle’s body measurements, such as chest girth, body length, and wither height, to predict body weight and achieved good results (Ozkaya & Bozkurt, 2009). In this study, the vision system will automatically capture the body measurements for body weight prediction. The RGB and 3D images will be combined to accurately extract each animal’s chest girth, hip-width, body length, and wither height. Models will be trained to associate the body measurements with body weight. As for broiler body weight, 3D images will be used to accurately extract each bird’s body volume, contour area, blob height, width, and length. Models such as PointNet and MLP will be combined and incorporate additional information such as age and breed information, to estimate body weight of each bird.

Elevated respiratory rate can indicate stress in animals, such as heat stress, the onset of farrowing events, and respiratory diseases. There has been no existing technology developed for continuous remote monitoring of animals’ respiratory rate. In this project, a video processing pipeline will be developed to extract respiratory rate and intensity of sows using 3D video processing techniques and AI tools for respiratory signal filtering. The developed technique will be able to be applied in other livestock sectors, such as cattle and chicken, for health and welfare monitoring.

 

Obj. 1b. AI tools for autonomous system perception, localization, manipulation, and planning for agroecosystems.

A. Introduction

Automation plays an important role in the entire agrifood supply chain. Even though traditional agricultural mechanics have been adopted in many agricultural production and processing applications, it is very challenging to implement robotics and autonomous systems in unstructured environments without advanced AI tools that can interpret the environment and make effective decisions required for localization, navigation, manipulation, and planning when coupled with sensor perception. Economic, agronomic, workforce, and environmental pressures, coupled with the need for complex robotic operations within limited equipment budgets, have resulted in a growing demand for AI-enabled robots. Advanced robotic platforms such as unmanned ground/aerial vehicles, multi-axis robotic arms, quadruped robots, soft robots and humanoid robots have expanded their application scope in the agrifood supply chain. New breeds of embedded processors and GPU-enabled servers provide the potential for Edge computational capacity that can run complex models without cloud service (due to limited internet access in farm fields) to support robotic field operations. AI techniques not only enhance the robotic sensing capabilities from multimodality sensors, but also boost the capability of robot control strategies via state-of-the-art robotic learning methods such as reinforcement learning and imitation learning to achieve human-level operation efficiency.

B. Detailed activities/procedures

Our research activities will be concentrated on applying AI tools that will improve robotic perception, localization, manipulation, and planning for agroecosystems. We will derive closed-loop systems based on large datasets that include both the multimodality information collected from the sensors attached to unmanned platforms, and the corresponding ideal control actions derived from human operators.  The closed-loop system’s goal is to measure, monitor, and control a process by monitoring its output, comparing it to the desired output to reduce the error. Generally, the data will be collected from not only from the field conditions but also the simulation or digital twin environments.  Different sensors (Lidar [2D & 3D], imaging, distance, rotary, GPS, IMU, tactile , etc.) will be used for the field trials and in the simulation. AI tools, including CNNs and transformers, are critical to be developed for effective feature extraction to reduce dimensionality from the multimodality sensor information and to use appropriate and the most relevant features as input for building the predictive models and improving prediction accuracy. The actuators in the system will include off-the-shelf electronics, devices, robotic manipulators or customized machinery and automation systems. The control algorithms will be developed, which include traditional control algorithms based on the system's error, and bringing the output back to the desired state. New data-driven control methods including reinforcement learning and imitation learning will also be developed, which can directly map the network-processed sensor features into human actions and optimal decisions.  These new methods will minimize the uncertainty associated with hard-coded control, and can better handle the system non-linearity in the real-world production and processing systems.

To develop and validate the predictive models and algorithms, we plan to explore, develop and implement approaches that will use various forms of edge processing either with systems like smart cameras or AI-enhanced embedded processors like Nvidia Jetson Nano or Xavier or with GPU enabled Edge servers on a local network in the field. We will conduct field trials at multiple environments and various geographical areas for their applicability. The production field experiments will be conducted over multiple years. Research on the use of different protocols will be conducted through best practices from ongoing AI projects from multistate members. Pertinent results will be disseminated through peer-reviewed publications and presentations at national and international conferences.

 

Obj. 1c. Natural resources scouting and monitoring.

A. Introduction

Soil is one of the foundational natural resources that support all terrestrial ecosystems. The soil ecosystem services include but are not limited to the provision of food, supply, and cycling of plant nutrients that support various land use and cover types, the regulation of the environmental functions through biodiversity, water purification, and climate resilience, and provide other services that support culture as the foundation of local landscapes. Rapid population growth and excessive soil use have resulted in soil degradation worldwide, including soil organic matter depletion, soil erosion, soil compaction, and salinization, which have caused loss of biodiversity, decreased agricultural productivity, water pollution, eutrophication, and greenhouse gas emissions. Concerns have been raised that soil health is being compromised in many parts of the world, and the degradation is becoming increasingly apparent across the U.S.

The function of soils in providing ecosystem services is enhanced by the formation of soil structure, which is an aggregate of organic-inorganic complexes formed from decomposition products of plant remains bound to soil minerals. Soil structure regulates water retention, drainage, and aeration essential for plant growth. In agriculture, the degradation of soil structure has been caused by the depletion of soil organic carbon, one of the important soil health indicators, due to various factors, such as the intensification of agricultural production, low adaptation of conservation practices, and a lack of understanding about pedogenesis and its relationship with agricultural productivity.

The potential of soils to sequester carbon and reduce anthropogenic CO2 emissions, thereby  mitigating climate change, has been widely recognized and has been an active research area for decades. Soil health indicators, other than soil organic carbon, have also been well studied through traditional methods, which involve intensive field and laboratory work that is expensive, time-consuming, and labor-intensive. Recent advancements in sensing and computational technologies have generated large volumes of soil and characterization databases, which can be leveraged to rapidly, cost-effectively, and accurately quantify soil health indicators. AI-based approaches are being explored to identify relationships in big data, facilitating real-world applications for soil health management and precision agriculture.

Water resource conservation is another critical global challenge. Nutrients (e.g., nitrogen and phosphorus) from agricultural runoff coupled with climate change (e.g., warmer temperatures and rainfall anomalies) have led to an increase in harmful algal blooms (HAB) worldwide. Efficiently monitoring water quality with high spatial and temporal resolutions remains a challenge. While remote sensing technology has been used to develop predictive models for water quality parameters that are difficult to measure intensively, not all water quality variables, such as off-flavors and toxins, cause changes in the spectral reflectance of surface water. The integration of spatiotemporal dynamics from multimodal remote sensing data, weather data, and geographic data can contribute to accurate predictions of the non-optically active water quality parameters and to forecasting of HAB outbreaks. Deep learning algorithms hold great potential to harness such integrated data, enabling predictive models that can improve the decision-making of water resources managers and policymakers.

B. Detailed activities/procedures

The work will begin with multistate data development with quality assurance so that diverse datasets can be combined and compared effectively. Existing and newly collected soil cores, laboratory measurements of physical, chemical, and biological properties, VNIR/MIR spectra, in/ex situ proximal and remote imagery, soil characteristics, water quality, weather, and terrain information, and management histories will be harmonized under common protocols. Shared sampling depths, bulk density procedures, consistent documentation of field and lab metadata and analyses, along with advanced computational modeling, will allow reproducible integration from plot to a larger scale.

AI models for indicators of soil, water, and plants will be developed using both traditional data‑driven and physics‑informed approaches. Training targets will include soil (e.g., SOCbulk density, texture, pH, cation exchange capacity), key nutrient pools such as nitrate, ammonium, and plant‑available phosphorus and potassium, while water‑quality related outputs will address nutrient export risk and bloom susceptibility within a watershed context. Topography, weather, and other relevant information will be considered to add if appropriate. Multimodal data fusion will incorporate spectra, imagery, terrain, land cover, management histories, soil dynamics, and hydrologic connectivity. Monitoring of bulk density and soil carbon stocks will occur through a coordinated network of sentinel sites spanning soil types, landscape positions, and management systems.

Model performance will be evaluated with an emphasis on uncertainty quantification and transferability across regions and seasons, employing techniques such as domain adaptation, hierarchical modeling, and ensemble methods to generalize beyond the calibration domain. Decision‑relevant outputs will be co‑designed with producers, conservation planners, and water managers to support timing, placement, and rate decisions that jointly benefit productivity and water quality.

Three‑dimensional computer vision will be applied to cores, monoliths, and soil pits using accessible workflows to generate meshes and voxels that capture aggregates, pore networks, root channels, and horizon geometry. Derived metrics such as aggregate size distributions, macropore connectivity, and root channel density will be related to infiltration proxies, bulk density, soil organic carbon fractions, and nutrient retention. The resulting models will be curated with interpretive overlays and educator guides, and they will be integrated into classrooms, field days, and digital exhibits, including VR and AR demonstrations, to improve understanding of structure–function linkages that control nutrient cycling, water movement, and carbon stabilization.

A publicly accessible web portal will allow users to upload mid‑infrared spectra and receive property estimates accompanied by uncertainty metrics, documentation, and guidance on appropriate use, thereby broadening access to spectral analysis in research and practice. This component incorporates and extends the originally proposed plan to obtain soil core samples, pair them with VNIR/MIR spectra and reference analyses, and deploy AI models that remain accurate when transferred across instruments.

Obj. 1d. Phenotyping and genotyping.

A. Introduction

In recent decades, plant genetics research has focused on developing crop varieties with enhanced traits such as high yield, environmental stress tolerance, and disease resistance (Cuenca et al., 2013; Rambla et al., 2014). Current breeding methods require many years to develop, select, and release new cultivars (Sahin-Cevik et al., 2012). New breeding methods, such as genomic selection, incorporate genomics, statistical and computational tools, and accelerate cultivar development (Vardi et al., 2008; Zheng et al., 2014; Albrecht et al., 2016). A key requirement for implementing these new breeding methods is creating a large and genetically diverse training population (Aleza et al., 2012) and genotyping datasets (Yoosefzadeh Najafabadi, 2023). Hence, large-scale experiments in plant phenotyping are critical because the accurate and rapid acquisition of phenotypic data is important for exploring the correlation between genomic and phenotypic information. Traditional sensing technologies for evaluating field phenotypes rely on manual sampling and are often labor-intensive and time-consuming, especially when covering large areas (Mahlein, 2016; Shakoor et al., 2017). Additionally, field surveys for weed and disease detection to create plant inventory and assess plant health status are expensive, labor-intensive, and time-consuming too (Luvisi et al., 2016; Cruz et al., 2017; Cruz et al., 2019). Small unmanned aerial vehicles (UAVs) equipped with various sensors have recently become flexible and cost-effective solutions for fast, precise, and non-destructive high-throughput phenotyping (Pajares et al., 2015; Singh et al., 2016).

UAVs allow growers to constantly monitor crop health status, estimate plant water needs, and even detect diseases (Abdullahi et al., 2015; Abdulridha et al., 2018; Abdulridha et al., 2019). They represent a low-cost method for image acquisition in high-resolution settings and have been increasingly studied for precision agricultural applications and high throughput phenotyping. UAVs and machine learning, an application of AI, have been increasingly used in remote sensing for genotype selection in breeding programs (Ampatzidis and Partel, 2019; Ampatzidis et al., 2019; Costa et al., 2021).

These methods have achieved dramatic improvements in many domains and have attracted considerable interest from both academic and industrial communities (LeCun et al., 2015). For example, deep convolutional neural networks (CNNs) are the most widely used deep learning approach for image recognition. These networks require a large amount of data to create hierarchical features to provide semantic information at the output (Cruz et al., 2017; Krizhevsky et al., 2012; Simonyan and Zisserman, 2015). With the increasing access to large amounts of aerial images from UAVs and satellites, CNNs can play an important role in processing all this data to obtain valuable information for breeding programs. Since labor shortage is a major issue, remote sensing and machine learning can simplify the surveying procedure, reduce labor costs, decrease data collection time, and produce critical and practical information for breeding programs.

Recent advancements in AI and remote sensing have further improved the efficiency and accuracy of phenotyping. Notably, the adoption of self-supervised learning (SSL) and foundation models such as Segment Anything Model (SAM) and Vision Transformers (ViTs) have allowed for more robust feature extraction with minimal labeled data. These models, along with Graph Neural Networks (GNNs) and multimodal AI frameworks, now enable seamless integration of genomic, phenotypic, and environmental data, revolutionizing breeding programs.

Over the past five years, we have successfully: Selected and standardized proximal sensors and data processing methodologies; Established partnerships with other universities to replicate trials where feasible; Developed a common database for training AI-based phenotyping algorithms; Deployed test plots for UAV-based data collection across multiple locations; Built algorithms and platforms for data upload, analysis, and visualization; Designed AI-driven algorithms to extract plant parameters and analyze genotype performance; Trained technical support personnel in remote sensing data collection and analysis.

While early breeding programs relied primarily on crop genotyping to characterize genetic architecture, modern breeding efforts now integrate multiple layers of biological information to better capture the molecular and physiological basis of complex traits. Current and future plant breeding efforts rely on the integration of high density molecular datasets (including genomic, transcriptomic, epigenomic, proteomics and metabolomic data) with phenotypic data, to capture the genetic architecture and physiological responses underlying the complex agronomic traits. The convergence of omics technologies with Machine Learning will enable breeders to efficiently analyze multidimensional data and accelerate cultivar development (Yoosefzadeh Najafabadi, 2023).    

In the next five years, we will focus on refining and improving the models developed in Years 1 to 5, incorporating cutting-edge AI advancements and optimizing our methodologies for enhanced efficiency and scalability.

B. Detailed activities/procedures

Building on the foundational work established in the first five years, our next-phase activities will focus on refining existing methodologies and incorporating recent advancements in AI and remote sensing.

We will enhance AI-based feature extraction by integrating self-supervised learning (SSL) and foundation models such as Segment Anything Model (SAM) and Vision Transformers (ViTs) to improve phenotypic trait identification with minimal labeled data. Additionally, Graph Neural Networks (GNNs) will be explored to enhance phenotypic trait predictions by leveraging structured data relationships.

To further improve genotype-phenotype associations, we will leverage multimodal AI approaches, integrating genotypic, phenotypic, and environmental data to refine selection criteria in breeding programs. The application of transformer-based models will enable enhanced pattern recognition, improving predictive analytics for genotype performance evaluation.

To overcome data limitations in disease detection and trait mapping, we will introduce synthetic data generation techniques, utilizing generative adversarial networks (GANs) and diffusion models to create augmented training datasets. This approach will improve AI model robustness and generalization across varying environmental conditions.

Another key focus will be the adoption of real-time Edge AI for in-field phenotyping. We will deploy onboard AI processing for UAVs to enable real-time classification and trait assessment, minimizing the need for extensive post-processing. The integration of edge computing solutions will further facilitate in-field decision-making. To further advance ground robotic systems for high-throughput phenotyping. Currently, only a few research labs have developed ground robotic systems for phenotyping, and their integration into breeding programs remains limited. This is largely due to the low efficiency of ground robotic systems compared to drone-based systems. Over the next five years, we will leverage advanced AI algorithms to accelerate the development of autonomous robotic systems and integrate it into the breeding programs for various crops.

In addition to technological advancements, we will continue refining existing algorithms for plant parameter extraction, improving accuracy in biomass estimation, disease detection, and stress evaluation. Expanding test plots across diverse agro-ecological zones will provide additional validation for these algorithms.

 

Objective 2: Data curation, management, accessibility, security, and ethics

Obj. 2a. Create open source agricultural datasets following FAIR (Findable, Accessible, Interoperable, Reusable) principles

A. Introduction

AI algorithms are capable of analyzing large volumes of diverse datasets to uncover complex patterns and relationships that are often challenging to capture through traditional methods, thereby enabling the development of data-driven solutions in agriculture. While some progress has been made in standardizing and centralizing datasets for specific tasks—such as crops/weeds imagery (Deng et al. 2024) and livestock datasets (Bhujel et al., 2025)—significant gaps remain in the availability of comprehensive, well-curated datasets that fully support AI development in agriculture. Effective coordination among land-grant universities will be essential to create standardized datasets in both crop and livestock production systems for the development of AI applications.

Over the last several years, the team has initiated and contributed to numerous AI-based efforts, resulting in the generation of various information-rich datasets. This Multistate project aims to integrate these datasets into a common, shareable, and standardized environment. Such an effort is essential to maximizing the value of these datasets and accelerating AI innovation across institutions. However, the preparation of high-quality datasets is nontrivial, requiring substantial investments in data acquisition, categorization, annotation, standardization, and secure handling. Additional challenges include data encryption, de-identification, compliance with data-sharing and privacy requirements, all of which can create significant barriers to effective reuse. 

Coordinated data sharing offers a powerful mechanism to overcome these challenges. Publicly available, standardized datasets can significantly reduce the duplication of efforts associated with data collection and curation, while enabling benchmarking, validation, and comparison of AI algorithms developed by different research groups. Within this framework, land-grant universities are well-positioned to serve as trusted gatekeepers of agricultural datasets, ensuring broad access to high-quality, multidimensional data from various sources. This role also allows participating institutions to share knowledge of successes and failures of AI-based methodologies, fostering transparency, reproducibility, and continuous methodological improvements. Developing these datasets also fosters interdisciplinary collaboration among computer engineers, data scientists, and statisticians, while also supporting industry and startups in utilizing the data to create AI solutions and deliver services to end-users.

B. Detailed activities/procedures

Our priority will be to launch an inquiry to various land-grant universities to identify publicly available databases that they maintain. We will extend that to private companies affiliated with those universities and more. This way, we will compile a list of participating units and their capabilities of sharing or exchanging information. A standardized survey/questionnaire will be sent to all unit heads of extension programs, and the relevant research will be identified. 

The group will also identify publicly available datasets and create an easy-to-search catalog. Our researchers will test the accessibility of USDA-established databases and identify databases that include crop phenotyping data, imagery data sets, crop and livestock production and management datasets, satellite imagery, soil information, weather data, and economic data. Multiple modalities will be included in those datasets, and they will be as de-identified as possible. Standardization will be the key element, and for that effort, a thorough literature review will be conducted to identify established practices in the field. We will be building upon standardizations with an emphasis on AI tools development. The questionnaire will also be sent to the various companies like Oracle, Microsoft (Azure), as well as Digital/precision Agriculture companies like Mothive, Ag-Analytics, Agri-Data, and more working in the field of data management that have expressed an interest in participating. 

After combining the feedback from the questionnaires above, the group will develop new publicly accessible, large-scale image datasets designated for agricultural vision tasks (e.g., weed detection and control) with image- and pixel-level annotations and will benchmark the state-of-the-art deep learning algorithms for the datasets. Many small-scale datasets exist at this stage in various universities, so bringing them together will be a challenge but also a straightforward process after a few affiliated universities create the first venues of collaboration. The main source of these datasets will be the initial core researchers in this Multistate proposal, as explained in Objective 1.

The group will create and share de-identified datasets from participating farmers, extension agents, and other partners. This will include obtaining new samples and laboratory data to create datasets. These automated processes are already in place, and some of the co-PIs have already been extending them (LSU’s connection with Ag Analytics is such a platform that is now being explored further). The group will share benchmark attempts and the corresponding standardized datasets. All de-identified information, including transfer protocols, de-identification protocols, and final results, will be available to participating universities. Finally, we propose the creation of various educational programs for students and extension agents, but also for interested researchers in the field, whose goal would be to provide a robust understanding of the theoretical underpinnings of the basic AI models, in an Agricultural setting. 

Yearly meetings and workshops on applications of AI with an emphasis on hands-on learning, especially for students in agriculture, as well as forums and yearly talks about the ethical issues in AI and the effect they may have on various stakeholders, will be established. The program will be connecting our experts to specialists in Computer Science, who will then participate in various workshops and presentations about the ethical issues with the use of AI and the better distribution of results.

 

Obj. 2b. Data Standardization and testbed development

A. Introduction

One of the major challenges hindering the widespread adoption of Artificial Intelligence in agriculture is the absence of standardized datasets and integrated software solutions tailored to handle agricultural big data for end-users. Agricultural data is inherently complex and heterogeneous, encompassing diverse sources. For example, in crop production, data is generated by different sensors at varying capacities and scales (e.g., soil moisture, satellite imagery, drone data, weather information, field notes, genetic profiles, and management records). Data generated from livestock production systems is also vast, diverse, and increasingly digital, encompassing a wide range of biological, environmental, and management variables. This includes real-time data from precision livestock farming technologies such as wearable sensors, RFID tags, automated feeders, climate control systems, and video monitoring, which track metrics like animal movement, weight, feed intake, temperature, reproductive status, and health indicators. In addition, forage and pasture management and quality data, genomic and phenotypic data, veterinary records, and milk or meat production outputs further contribute to the data system. In both crop and animal production systems, datasets often vary in format, structure, resolution, and semantics, making it difficult to aggregate, share, or apply AI models across different contexts. Thus, there is an increasing need to develop data standardization protocols based on FAIR data principles, allowing for the development and utilization of advanced ML/AI models. 

Besides the integration and standardization of the resulting datasets, the creation of dedicated testbeds presents an opportunity for a continuous data-generating process that will serve as a fixed point in the data creation and accumulation. A testbed is a platform for conducting rigorous, transparent, and replicable testing of scientific theories, computational tools, and new technologies.  We suggest adopting a digital portal or data hub with a user-friendly interface to integrate tools, implement algorithms, and facilitate data sharing and collaboration. For example, Texas A&M AgriLife developed a data portal named UASHub to facilitate UAS-based data communications. The UASHub integrates electronic field notes, raw and post-processed UAS imagery data for sharing, visualization, analysis, and interpretation of large volumes of UAS-derived data. The UASHub includes tools to integrate online data management and access tools, allowing research scientists to download both raw and processed geospatial data products to their workstations for further analysis. Establishing a collaborative, cloud-based, open-access platform that connects multiple experimental and data collection sites—and transforming this into an AI-ready testbed—will facilitate innovation, experimentation, testing, education, and community involvement in AI. By incorporating datasets from a range of production environments and management systems, the testbed will promote interdisciplinary research, teaching, and outreach across fields such as agriculture, animal science, geospatial science, computer science, data science, and statistics.

B. Detailed activities/procedures

Developing standardized datasets and AI-ready testbeds in agriculture requires a coordinated series of activities that span technical, organizational, and community dimensions. Based on the results obtained from assessing the publicly available databases (Objective 2A), we plan to evaluate the data quality of those databases and verify the quality of both the data and metadata. Additionally, rigorous planning and experiments will be conducted to develop data standardization protocols along with the pipeline to store and share metadata. Simultaneously, efforts will be made to design and deploy data integration pipelines that can clean, align, and aggregate diverse datasets such as sensor readings, satellite imagery, genetic information, and field observations. AI-ready testbeds will then be established by identifying and equipping representative production sites with real-time sensing technologies and cloud-connected infrastructure, enabling continuous data collection, transfer, and access. Annotation tools and quality control procedures will be implemented to label datasets for machine learning, facilitating the development and benchmarking of AI models. To ensure accessibility, a cloud-based digital platform or data portal will be created, offering user-friendly tools for visualization, analysis, and download of raw and processed datasets—similar to the UASHub developed by Texas A&M AgriLife. Education and outreach will be integrated through training programs, curriculum development, and hackathons to build capacity and foster interdisciplinary collaboration.

 

Obj. 2c. Agricultural data governance, ethics, and privacy

A. Introduction

As agriculture becomes increasingly data-driven, the need for robust data governance frameworks that address ethics and privacy is more critical than ever. Agricultural data is often collected from farms using a range of technologies such as sensors, drones, satellite imagery, and IoT devices, and may include sensitive information about land use, production practices, environmental conditions, and even personal data about farmers and workers. Without clear policies and protections, this information can be misused, leading to concerns around data ownership, consent, transparency, and equity. Our idea to map the data supply chain from Objective 1 will help us understand the data flow from its source through various intermediaries to its final end-users, which will provide a foundation for the economic, legal, and ethical implications of agricultural data.

Traditional open data licenses are often inadequate for the agricultural data value chain due to their narrow scope and failure to address critical stakeholder concerns. Licenses like CC BY-SA (Creative Commons Attribution-ShareAlike), ODbL (Open Database License), and GODAN (Global Open Data for Agriculture & Nutrition) typically require derivative datasets to be shared under the same terms but exclude other data derivatives—such as analytical reports, proprietary databases, or predictive models—which are often the most commercially valuable components. This omission creates ethical risks, exemplified by technology companies profiting from software built with farmer-shared data without providing recognition or compensation, potentially disincentivizing participation. 

This complexity is compounded by the diverse and often conflicting perspectives of stakeholders across the expanded data life cycle. Federal agencies may suggest open access for publicly funded research data, while farmers, unaware of these policies, may view the same data as private property. Similarly, early publication of research data related to a new crop variety could undermine a breeder's patent claims and commercial return. Agricultural Technology Providers (ATPs) also hold disparate perspectives in rights of data with some startups recognizing farmers’ ownership while well-established tech giants taking more aggressive control by legal leverages. Such conflicts can be evidenced by the "right to repair" lawsuit against John Deere regarding its smart tractors. The company placed restrictions on farmers’ access to the underlying data and ability to repair their equipment, asserting ownership over the software and data. This forced farmers to rely on authorized dealers for repairs, raising costs and limiting autonomy. Several lawsuits have been filed to force John Deere to change this policy, but the issue remains unresolved. These disparities highlight that a one-size-fits-all license is insufficient.

Therefore, a more nuanced approach is essential. This begins with a comprehensive analysis of existing governance documents and stakeholder interests to identify potential conflicts. The system must then develop staged, flexible frameworks that range from fully open data (including commercial derivatives) to partially open tiers that may reserve commercial rights or require profit-sharing. Ultimately, the goal is to create actionable, ethically balanced agreements, crafted with legal expertise, that clearly specify rights, obligations, and benefits for each stakeholder, ensuring the system remains both accessible and equitable.

B. Detailed activities/procedures

The path of establishing the proposed open data system begins with a comprehensive discovery phase: 1. aggregating disparate data governance policies from all relevant federal agencies, 2. soliciting and analyzing industrial contracts from ATPs who involved in research projects by land grant universities for a preliminary understanding of industrial practices of data governance regarding various devices, services, and farming operations, and 3) conducting in-depth social research (e.g., surveys, interviews) to capture the nuanced perspectives, concerns, and incentives of farmers, researchers, and ATPs. This intelligence is then used to construct a map of the entire agricultural data value chain, explicitly delineating each stakeholder's role, responsibilities, data inputs/outputs, and economic interests. This mapping is not merely descriptive but analytical, serving to identify and diagnose points of potential conflict in data ownership, usage rights, and value distribution. The system then facilitates a structured stakeholder engagement process to negotiate compromises and co-create mutually acceptable data-sharing agreements. Consensus-driven principles will be created during this process to direct the development of robust, standardized template agreements and flexible, staged data protocols, ranging from fully open data for commercial use to restricted models allowing only non-commercial derivatives or requiring profit-sharing. These processes will be summarized into educational materials to be integrated into courses in educational institutes (e.g., land grant universities). Ultimately, the system provides dynamic guidance and tools, empowering future participants to navigate this complex landscape, anticipate challenges, and reliably build upon a sustainable and equitable open data ecosystem.

 

Objective 3: AI adoption (technology transfer) and workforce development

A. Introduction

The human and social dimensions of AI knowledge transfer are essential for applying systems approaches to stakeholder engagement, science communication, and experiential learning to strengthen agricultural workforce development (Daigh et al., 2024). Research has shown that knowledge alone is often insufficient to drive adoption, especially when political, economic, or social factors play a role (Knowles, 1984; Kolb et al., 2014). The adoption of AI in agriculture is influenced by individual personality, information dissemination methodology, environmental conditions, structural factors, technology attributes, demographic variables, farmer education, household size, land tenure, access to credit, and the availability of extension services (Ruzzante et al., 2021). Facilitators of adoption and the establishment of trust between farmers and technology providers are especially critical for its successful implementation (Sood et al., 2022). Active collaboration between scientists and stakeholders can improve technology transfer and adoption. For example, field days that include active engagement of extension personnel and other stakeholders (e.g., growers) have proven to be effective tools in promoting experiential learning (Knowles, 1984; Kolb et al., 2014), which increase interest and willingness in adopting new technology (Rogers, 2003).

The major focus areas under this objective are to (i) support AI algorithms development team during the development of user-friendly digital tools/platforms by engaging stakeholders through User-Centered Design (UCD) (Parker, 1999) process (ii) train consultants, extension specialists, county agents, producers and allied industry on the use of digital and AI-based tools/platforms for precision farm management (technology transfer), through field days, workshop and outreach programs that provides hands on practical training, (iii) grow the number of next-generation experts on AI and digital agriculture tools development and use. The detailed procedure is described below to illustrate the method that will be followed to achieve this objective. We believe that this model will serve as a foundation for AI technology transfer, adoption, and workforce development. The experience and methods will be transferable and scalable to other states according to their needs and resource availability.

B. Detailed activities/procedures

Advancing Equitable AI in Agriculture Through User-Centered Design, Stakeholder Engagement, and Policy Advocacy

To successfully develop and deploy AI tools in agriculture, it is essential to go beyond merely sharing information and instead actively involve stakeholders at every stage of the process. This approach includes conducting structured needs assessments and involving end users from the very beginning of technology development, following a User-Centered Design (UCD) framework (Parker, 1999). User-Centered Design is an iterative design process that involves end-users in all phases of product development and addresses their needs (Barnum et al., 2020). Producer heterogeneity differences in farm size, technological readiness, educational attainment, and resource access play a substantial role in shaping adoption dynamics, which makes user-centred approaches imperative (Schimmelpfennig, 2016; Paudel et al., 2021). Policy briefs and evidence-based advocacy materials are developed to influence state and federal investments in digital infrastructure and to promote supportive mechanisms for integrating AI into agriculture. The integration of AI with IoT can optimize resource management, improve productivity, and reduce greenhouse gas emissions in agriculture, although policy interventions will be necessary to encourage broad adoption (Mohamed et al., 2021; Morkunas et al., 2024). This comprehensive approach integrates user-centered design, experiential learning, inclusive communication, modular education, and long-term impact assessment to ensure that agricultural advancements are technologically sound, socially grounded, and equitably accessible.

Empowering Agricultural Communities Through Extension Education and Technology Transfer

Historically, extension and outreach programs have often followed a deficit model, which operates on the assumption that people will adopt emerging technologies or scientific innovations if they are simply provided with more information (Nisbet & Scheufele, 2009). This project proposes a comprehensive extension, outreach, and education model that will function as both a national resource and a scalable framework for AI technology transfer in agriculture. The model will leverage a network of extension specialists and county agents trained in the application of AI tools for precision agriculture. For example, Texas A&M AgriLife Extension operates a broad network of Agriculture and Natural Resources Extension Agents and Integrated Pest Management (IPM) specialists across nearly every county. Virtual reality simulations will be used to offer immersive, experiential training on AI applications such as pest detection, irrigation management, and yield forecasting. Building upon this existing infrastructure, extension crop specialists in strategic locations will collaborate with county agents and IPM staff to facilitate the dissemination, evaluation, and refinement of AI-based tools across diverse cropping systems and ecological zones. To reach underserved and remote regions, mobile training units will provide in-person demonstrations in areas with limited digital connectivity. Social media platforms will also be used to share concise and engaging educational content that connects with younger and more technologically adept farmers. Community-Based Participatory Research (CBPR) initiatives can enable producers, researchers, and industry partners to co-design and refine technologies based on practical needs.

Additional strategies to deepen engagement include establishing regional Farmer-Led Innovation Councils to guide AI deployment, organizing participatory scenario planning workshops to explore potential future agricultural scenarios, and creating farmer innovation labs to serve as hubs for demonstrations, pilot testing, and user feedback. More collaborative approaches, such as active engagement between scientists and stakeholders, have demonstrated greater success in promoting technology transfer. For example, field days that include hands-on participation from extension agents, producers, and other stakeholders can effectively promote experiential learning and encourage the adoption of innovations (Rogers, 2003). Capacity building will be achieved through initiatives such as AI Literacy Bootcamps, which will provide short and intensive training for farmers and extension agents to demystify AI concepts and applications. Train-the-Trainer programs will equip local educators and extension agents with the skills and pedagogical tools needed to effectively teach AI concepts in rural communities.

Educational Innovation and Inclusivity in Agricultural AI: A Participatory Framework for Experiential Learning and Industry Collaboration

This project will develop modular, project-based curricula for undergraduate and graduate students, focusing on three core areas: high-quality data collection, data processing and analysis, and AI applications in agriculture. Partnerships with land-grant universities, minority-serving institutions, and agricultural technology companies will ensure diversity, inclusion, and geographic reach. AI Mentorship Networks will pair students and early-career professionals with experienced researchers and industry leaders to provide ongoing guidance and career development opportunities. Educational expansion will include dual enrolment AI-in-agriculture courses for high school students, certificate programs at community colleges for non-traditional learners, and gamified learning platforms that deliver interactive modules to increase engagement and retention. Micro-credentials or stackable certificates will be offered in areas such as data science, machine learning, and AI-driven farm management. To enhance inclusivity and accessibility, culturally responsive curricula will be designed to incorporate local agricultural practices and cultural contexts, making them more relevant to diverse audiences. Bilingual community ambassadors will act as liaisons and trainers in underserved regions. Low-bandwidth and offline-compatible AI tools will be developed to address connectivity limitations in rural areas. Digital literacy training will be delivered alongside AI modules to promote equitable access. Students will have opportunities for internships and apprenticeships with agricultural technology startups and research farms. Activities such as hackathons, innovation challenges, and interdisciplinary design sprints will encourage creativity and problem-solving skills. A new course on the human dimensions of AI will address the ethical, cultural, and socio-economic aspects of technology adoption, ensuring that graduates are not only technically skilled but also socially and ethically informed.

Monitoring, Evaluation, and Impact Assessment for Scalable Agricultural AI

The monitoring and evaluation framework will measure behavioral change metrics that capture shifts in farmers’ decision-making, risk tolerance, and technology adoption after interventions. Interactive dashboards will be developed to allow stakeholders and project leads to track tool adoption, training participation, and farm-level outcomes in real time. Longitudinal impact studies will assess environmental, economic, and behavioral changes resulting from AI adoption, which will inform the scaling and refinement of extension approaches.

The project will establish an open-source repository of datasets, use cases, and AI models to be shared nationally and globally. Collaborations with international organizations such as CGIAR and FAO, as well as global universities, will enable the adaptation of solutions for use in developing agricultural systems.

Measurement of Progress and Results

Outputs

• Obj. 1a: Major findings will be published, demonstrating the application of AI for crops, livestock, and postharvest systems. Trained deep learning algorithms will be available for crop yield estimation, crop stress assessment and management, monitoring of animal health, welfare, and behavior, and postharvest quality evaluation. Algorithms will be developed to automate the workflow of multi-scale data analysis, information extraction, information scaling, and synthesis for generating plots, field crop characteristics maps, or animal health and productivity indicators, supporting data-driven decision-making
• Obj. 1b: New AI methods will be developed for: (1) advanced perception, localization, and manipulation for robotic production (harvesting, management, pruning, etc.) and processing (sorting, handling, packaging, etc.) tasks, (2) object detection and dynamic mapping in the field and processing plants, (3) end-to-end path planning and obstacle avoidance approaches in agricultural settings and (4) edge computing approaches for perception, localization, and task planning.
• Obj. 1c: Soil cores, plant tissues, and water samples will be collected from multiple U.S. states and analyzed using VNIR and/or MIR spectroscopy alongside other sensing technologies and laboratory measurements of key physical, chemical, and biological soil indicators. These datasets will form the foundation for the development of AI‑based predictive models capable of estimating soil health properties directly from spectral inputs. Methods for calibration transfer will be created and tested for model robustness and accuracy. A comprehensive, openly accessible soil-water-plant database will be compiled to support modeling and methodological innovation. Using these resources, a web‑based prediction interface will be designed that allows users to upload lab and sensing data related to soil health and water quality. This online system will serve farmers, researchers, conservation agencies, and land managers by providing scalable, rapid assessments derived from advanced AI/ML approaches. The combined outputs—soil-plant-water datasets, physics (or soil science)-informed calibration models, transfer algorithms, spectral libraries, and decision‑support tools—will comprise a nationally relevant resource for soil health monitoring and natural resource stewardship.
• Obj. 1d: A standardized, multi-scale, multi-modal dataset including imagery and measurements of plant architecture, yield, yield-related traits, and disease across environments and time points. AI-based image analysis tools will be developed for predicting yield, drought response, and disease. Prototype AI-enabled prototyping platforms with integrated data visualization and decision-support tools will be developed to support variety selection and high-throughput phenotyping. A harmonized high-throughput multi-omic pipeline and centralized database will enable cross-location data interoperability and ML-driven crop improvement.
• Obj. 2a: A shareable database schema with specific access capabilities will be created as well as a methodology to connect to it safely with both adding and downloading capabilities. Recurring workshops discussing the theoretical underpinnings of AI models in agriculture as well as the ethical use of AI and various security issues.
• Obj. 2b: The primary end product from this objective will be a database management system: (1) a UAS-based platform to collect detailed and high quality HTP data for research plots or commercial fields, (2) automated procedures for data processing, analysis, and growth parameter extraction, to analyze, visualize and interpret collected data, and (3) web-based algorithms for data management and communication for all project scientists
• Obj. 2c: 1) A map of agricultural data value chain and specific logic of data governance (e.g., ownership of data and data derivatives) will be created by analyzing existing contractual documents in technology adoption cases in agriculture and the policy/regulations from government agencies. 2) A database of de-identified interview transcripts will be created from in-depth social studies targeting agriculture stakeholders, including farmers, researchers, and ATPs. 3) A data governance agreement template will be created based on previous results in facilitating different governance scenarios.
• Obj. 3: Several annual workshops will be developed on a need basis. In the last five years, several of these specialized workshops were provided during the annual AI in Agriculture Conference that includes hands-on experience on software use, data analytics, etc. Also, technology expo, seminars, webinars, podcast series, field days, and in-service training will be established that target primarily extension agents, farmers, and allied industries, focusing on the uses of AI in agriculture. Extension publications, infographics, and fact sheets will be developed to present advancements in AI-based technologies. Classes with a focus on AI applications in agriculture and natural resources will be added and evaluated on a yearly basis. Certifications and Specialization programs will be added to the already established Minors in relevant fields.

Outcomes or Projected Impacts

• Obj. 1a: The projected outcomes include improved preparation for harvest and storage through enhanced yield prediction, timely implementation of management strategies to minimize the risk of pest invasion, and reduced losses associated with low-quality produce entering the supply chain. These advancements are expected to lead to more efficient resource use and higher overall profitability. Farmers and stakeholders will benefit from the deployment of advanced algorithms for accurate yield forecasting and variable rate input applications, optimizing operational decisions. In the area of animal health and welfare, expected impacts include the development of automated assessment tools that support more effective farm management, improved animal welfare, increased productivity, and enhanced profitability across livestock operations.
• Obj. 1b: Automated bioproduct detection and processing systems using AI to improve performance and efficiency will enhance site-specific crop/animal management and processing practices to increase yield, reduce cost, enhance biosecurity, and improve grower profit. Implementation of AI-based edge processing will improve execution speed and perception effectiveness in a lower cost embedded vision controllers.
• Obj. 1c: The AI‑calibrated models and the web‑based soil prediction portal developed under Objective 1c will expand the accessibility and accuracy of soil health assessments for diverse end users, including farmers, county agents, researchers, and educators. By reducing dependence on traditional laboratory methods, these tools enable more frequent, lower‑cost soil monitoring that directly supports improved nutrient management, soil conservation planning, and precision agriculture. The integration of computer vision and 3D soil modeling will also enhance educational and outreach impacts. High‑resolution, spatial–temporal reconstructions of soil profiles, landscapes, and natural‑resource conditions will help students better grasp the complexity of agroecosystems, including spatial variability, soil formation processes, and linkages between soil, water, and plant interactions. The development of virtual 3D soil monoliths, soil pits, and field environments, accessible through virtual reality (VR), will allow learners to explore realistic field scenarios regardless of weather conditions, resource availability, mobility limitations, or geographic constraints. These immersive, computer‑vision–based tools will sustainably improve experiential learning for students who may not have access to field sites, instructors, or higher‑education resources. Through virtual soil pits, cross‑sections, and field tours, learners can engage directly with soil morphology, texture, structure, horizons, and other diagnostic features traditionally observed only during in‑person fieldwork. This democratizes access to hands‑on environmental education and strengthens training in environmental and agricultural sciences. Collectively, these outcomes will improve the scientific literacy of future agricultural professionals, promote broader participation in soil and environmental sciences, and support the development of a digitally skilled workforce capable of applying AI, sensing, and computer vision tools. Long‑term impacts extend to enhanced soil stewardship, reduced off‑farm nutrient losses, increased agroecosystem resilience, and improved environmental sustainability across working lands.
• Obj. 1d: New knowledge will be gained on efficiently applying state-of-the-art deep learning models, pretrained on large-scale color image datasets, to crop phenotyping tasks with limited labeled data and heterogeneous sensor modalities. The work will foster transdisciplinary, multi-institutional collaboration in high-throughput phenotyping (HTP), supporting breeders and agricultural scientists in elite genotype selection and interpretation of experimental treatments. Adoption of standardized multi-omic integration pipelines will improve data consistency and interoperability across different breeding programs, enabling more accurate identification of elite genotypes and trait-associated markers. This will reduce the data generation costs by facilitating data sharing and reducing duplication efforts while accelerating genotype selection decisions.
• Obj. 2a: AI-based algorithms depend heavily on the existence of large, clean, and information-rich databases. By combining multiple datasets, the users will see a dramatic increase in their algorithms’ predictive percentages and pattern recognition ability.
• Obj. 2b: This objective will enable transdisciplinary scientists to communicate and exchange information and accelerate the development of agricultural applications for crop management. The data portal has the potential to deliver tools and methodologies for UAS-based HTP, enabling the development of AI-based cognitive tools. Data gathered using the proposed framework would provide users with a high level of both spatial and temporal details of crops at a scale.
• Obj. 2c: The key logics of data governance in the value chain will be clear to all stakeholders, mitigating underlying misunderstandings and ambiguities. Such clarity will bring ethical and stable relationships among different stakeholders in adopting data-intensive technologies with complex business structures. Potential conflicts and related costs can be significantly reduced by the developed agreement, which may provide a principled framework to address these conflicts. Most importantly, all stakeholders will obtain an opportunity in learning the related knowledge and best practices in adopting new technologies.
• • Obj. 3: An outcome of this Objective is to increase the acceptability, awareness, and trust of AI by stakeholders in the agriculture industry while building workforce capacity to meet emerging demands for AI-enabled food production and processing systems. This project will serve as a nexus point for education, training, and outreach in AI for agriculture and increase knowledge of available careers in AI for underrepresented groups, including minorities, women, rural residents, and other disadvantaged populations. Every year, we plan to publish over 200 extension articles, release over 100 podcasts, publish over 120 peer-reviewed articles, organize at least 10 workshops during the calendar year, and host AI in Ag. Conference, and educate more than 5000 students, post-doctoral researchers, faculty, farmers, data engineers, agricultural and biosystems engineers, and other stakeholders on AI developments for precision agriculture applications, nondestructive testing methods, large language applications in information dissemination, etc. via our education, extension, and outreach activities

Milestones

(2027):Obj 1: In year 1, multiple multistate research teams will be built within existing members and newly joined members to cover different tasks (a) AI for food, crop, and animal production and processing, (b) AI for robotics, (c) AI for natural resources, and (d) AI for phenomics and genomics. Obj. 2b: During year 1, we will select and develop testbed fields at various locations, and also a cloud-based “Data Portal” and software to upload, analyze, and visualize remote sensing data. Various educational workshops will be organized to train users of the portal and get feedback. Obj. 2c: We will develop an automated process to solicit government policies/regulations regarding data governance and industrial contractual documents from the internet. A natural language processing (NLP) packet will be created to extract the key logic and knowledge from these documents. A social study questionnaire will be developed based on the extracted logic and knowledge so that the following interviews or surveys will be grounded in specific cases. Obj. 3 A major milestone expected every year in the next five years is the annual workshops organized along with AI in Ag. Conference. Currently, there are more than five stations interested in hosting the annual AI conference beyond the time of this project, along with workshops that focus on hands-on training in data analytics for all stakeholders from academia and the agrifood industries. Also, stations will work together to consider proposal ideas that target grants such as the USDA Strengthening Agricultural Systems (SAS) for Artificial Intelligence (AI) for K12 Food and Agricultural Sciences. A working group formed for the extension and outreach programming among stations involved in extension activities will work to regularly share information on best practices to enhance various outreach programs such as webinars, seminars, field days, in-service training, and extension publications.

(2028):Obj 1: Standardized data acquisition and AI adoption protocols will be discussed for different applications, including (1a) yield monitoring, pest infestations, postharvest quality evaluation, animal behaviors, (c) soil properties, (d) crop phenomics and genomics. Obj. 2a: Within the first two years of the project, a common repository will be formed, and the necessary transfer protocols, safety protocols, and access protocols will be established. We will continue collecting data, expand testbed locations, add more users, and improve the analytical and visualization algorithms. Obj. 2c: We will continue and expand the educational workshops on uses of the system and receive feedback from participants. The map of data governance logic and practices will be created from solicited documents. We will start social studies to learn about stakeholders’ perspectives regarding data governance. Obj. 3 Sponsor and organize AI in Ag. conference and related workshops. Submission of SAS proposal for K12 stakeholder training in AI. The working group on extension and outreach programming continues to meet to share ideas and develop an online multi-institutional program that targets consultants and service providers on how to include AI tools in their portfolio.

(2029):Obj 1: There will be at least one collaborative research project and one submission of multi-university research proposals for each individual research team. The AI model's robustness, reproducibility, and transferability will be discussed and studied. Obj. 2b: In years 3-5, we will expand and refine the activities listed in year 2 and add many surveys to determine the use of the resulting portal and extend its capabilities. Obj. 2c: The social studies will be completed, and the results will be analyzed to reflect stakeholders’ perspectives about data governance. Obj. 3 Sponsor and organize AI in Ag. conference and related workshops. If funded, involved stations will execute the SAS project that provides training to teachers and students in K12 on AI applications in agrifood systems. In year 3, stations will begin to work on developing LLM tools (chatbots, Agentic AI, and Autonomous AI tools) that can enhance information dissemination in agrifood production in a way that is timeless and asynchronous extension with limited interference of experts based on tools such as Agentic AI know-how of AI applications. The working group on extension and outreach programming continues to share ideas and organize joint training sessions for stakeholders.

(2030):Obj 1: Items listed in Year 2 will be refined and improved. The database from different research groups will be organized and structured, and the AI model developed will be organized and disseminated. Obj. 2a. Connection with existing databases will be established on the 4th year mark. Obj. 2b: In years 3-5, we will expand and refine the activities listed in year 2, and add many surveys to determine the use of the resulting portal and extend its capabilities. Obj. 2c: An agreement of data governance will be drafted to facilitate all data value chains in this project, with consideration of ethical and fair collaboration and data management. All collected information and derived results will be integrated into educational materials. Obj. 3 Sponsor and organize AI in Ag. conference and related workshops. If funded, the SAS outreach-training project execution will continue. Also, the LLM tools will be tested in-house across different stations. The working group on extension and outreach programming continues to share ideas and organize joint training sessions for stakeholders.

(2031):Obj. 1a: At the end of the five-year project, key milestones will include the development of AI models and tools with high accuracy for predicting yield, detecting pest infestations, and assessing the quality of both in-field and postharvest agricultural products. A structured and comprehensive database, along with standardized data acquisition protocols, will be established as a foundational step to enable effective application of AI approaches for crop health and pest management. For animal production, milestones include (1) development of a generalized computer vision system to estimate cattle & broiler bodyweight, evaluate broiler behaviors, and remotely access the respiratory rate of sows, (2) commercialization and deployment of the smart monitoring systems in livestock systems and non-destructive models that include reconstructed HSI data-RGB-based for food contaminant source detection and quantification. Obj. 1b: At the end of the five-year project, key milestones will include the development of AI models and tools for improving robotic-enabled crop and animal production and processing practices. Structured simulation platform, standardized data processing protocols, and a robotic control pipeline will be established to improve 1) specialty crop harvesting, 2) automated greenhouse management, 3) automated postharvest management, and 4) automated animal production house management. 5) automated food processing. Obj. 1c: In years 4-5, a harmonized spectral library containing soil health measurements and corresponding spectra will be finalized and documented. Using the validated models and transfer algorithms, a user‑friendly, web‑based predictive portal will be developed, tested with stakeholders, and released publicly. Documentation, tutorials, and training materials will also be created to support adoption by farmers, agencies, and researchers. Obj. 1d: Enable the training and deployment of multimodal machine learning models for genomic selection, trait prediction, and genotype–environment interaction analysis in subsequent spin-off projects. Completion of an integrated genotyping and multiomics platform for large-scale AI model refinement, real-time decision-support development, and cross-location predictive deployment. Obj. 2a: The dissemination of information will begin in earnest in year five. Obj. 2b: In years 3-5, we will expand and refine the activities listed in year 2, and add many surveys to determine the use of the resulting portal and extend its capabilities. Obj. 2c: We will use the agreements to guide the activities for all involved parties to enhance the trust of the partnership with all stakeholders. The developed educational materials will be used in extension events and other related activities to leverage stakeholders’ awareness and knowledge about data governance. Obj. 3. Sponsor and organize AI in Ag. conference and related workshops. USDA SAS project execution will continue. At the end of year 3 of the SAS project (year 5 of the multistate proposal), we expect to have trained 2000 stakeholders across several stations that will participate in the project. The LLM chatbot that will be developed will be tested with farmers, and feedback will be evaluated to improve users' experience. The working group on extension and outreach programming continues to share ideas and organize joint training sessions for stakeholders.

Projected Participation

View Appendix E: Participation

Outreach Plan

This project will train county Extension agents, crop consultants, producers, and allied industries in the practical use of AI-based tools for farm decision-making. Training Extension personnel is expected to have a high multiplier effect, as county agents are a primary source of trusted, science-based information for producers on agronomy, disease, and pest management (Hall et al., 2003).

Texas A&M AgriLife Extension provides an illustrative example, with agriculture and natural resources County Extension Agents (CEAs) and Integrated Pest Management (IPM) agents serving 250 of Texas’s 254 counties. Given their critical role in supporting producers, these agents must remain current on emerging technologies and management options. Accordingly, this project will establish a coordinated network of Extension crop specialists at three strategic regions of Texas: the High Plains (Lubbock–Amarillo–Vernon), Central Texas (Temple, College Station), and South Texas (Wharton, Corpus Christi, Weslaco). Specialists from each region will train and engage approximately 75 CEAs and IPM agents in the evaluation, adaptation, and dissemination of AI-based crop management tools developed through this project. Additional stakeholders will include crop consultants, agricultural industry partners, and producers.

Extension and outreach activities will include: (i) development of digital education and communication materials; (ii) a podcast series highlighting AI-based agricultural technologies; (iii) topical webinars on emerging developments; (iv) in-service training for Extension agents (both in person and via the eXtension network); (v) field days; (vi) technology showcases and expos; and (vii) Extension publications, infographics, and fact sheets.

AI in Ag. Conference: The annual conference in AI unites leading researchers, industry innovators, and federal partners to shape the future of agriculture systems through AI-driven agricultural innovation. It is a major outreach outlet for our multistate group. During the conference, there are several workshops that are organized on hands-on applications, field days to companies that specialize in AI.

Organization/Governance

This multistate research project will be administered by three elected officers: a Chair, Vice Chair, and Secretary, who together will constitute the Project Executive Committee. The Executive Committee will oversee project activities, coordinate efforts among participating institutions, and facilitate annual project meetings.

Officers will be elected for two-year terms to ensure leadership continuity. At project initiation, all officers will be elected, with terms concluding at the end of the second annual meeting. Upon completion of each term, the Vice Chair will assume the role of Chair, the Secretary will advance to Vice Chair, and a new Secretary will be elected by the project members.

Annual meetings will be held to share project progress, exchange findings, and foster collaboration among participants. Administrative guidance will be provided by an assigned Administrative Advisor and a NIFA Representative.

Literature Cited

Attachments

Land Grant Participating States/Institutions

Non Land Grant Participating States/Institutions