Using AI to Increase Yield and Disease Resistance in Oats

Using AI to Increase Yield and Disease Resistance in Oats

ARS is confronting the biggest challenges to staple crops like grains. Craig Carlson is a Research Geneticist in the Cereal Crops Improvement Research Unit in Fargo, ND. His work focuses on oat quantitative genetics, with an emphasis on developing high-yielding varieties that have superior grain quality and stable disease resistance.

Welcome Dr. Carlson to Under the Microscope

UM: What are the main area(s) of focus in your research on oats?

CC: My lab is focused on increasing genetic gain in oats, which simply means making steady and measurable improvements over time. These improvements can include things like higher grain yield, better disease resistance, nutrition and quality, etc. Genetic gain happens when breeders select the best plants based on these traits and use them as parents to create the next generation. If selection is effective, the average performance of the breeder's population keeps improving from one generation to the next. It's like picking the best players each season - if you keep picking top performers, then your team (breeding program) will most likely get better year after year. 

UM: How has agricultural research historically addressed oat genetics?

CC: The most devastating disease of wild and cultivated oat is crown rust, which is caused by a highly volatile fungal pathogen. In fact, oat crown rust is so aggressive that resistance genes deployed in new oat varieties are often overcome within 5 years, which is exceptionally frustrating for scientists and growers, alike. However, there is a big push from stakeholders to identify and deploy other forms of resistance that are more stable or not as vulnerable to breaking down, moving from single‑gene "boom‑and‑bust" resistance to quantitative, multi‑gene strategies. 

The traditional method for rating crown rust disease, called the Cobb method, dates to the early 20th century, and is still used by the USDA. However, this scale tells us only about the level of disease on the host plant. This is where our work comes in: To provide a mechanistic understanding of plant-pathogen interactions by quantifying the actual signs and symptoms of disease using an offshoot of AI called "computer vision" to help identify more durable sources of disease resistance. 

Dr. Craig Carlson in the field

UM: Are there other issues that are important to understand about oats?

CC: Absolutely. Oats are grown on only ~1.5 million acres in the U.S., yet the global market is worth about $9 billion. That demand is rising because oats are naturally gluten‑free and proven heart-healthy, but the crop has lagged in research investment compared to higher-value commodity crops like wheat and corn. Beyond diseases, we tackle stress tolerance (heat and drought), standability, and end‑use quality, all of which influence whether growers will risk planting more oat acreage. Limited resources mean breeders often must prioritize rust resistance, yet our multi‑trait genomic tools allow us to track yield and quality traits at the same time, so we don’t rob Peter to pay Paul. 

UM: What does/could AI enable you to do that was difficult or impossible without it?

CC: AI enables us to rapidly and accurately measure signs and symptoms of diseases. Our deep‑learning model recognizes pustules (oblong orange-yellow lesions on leaves containing rust spores) even under variable light conditions, something a human can’t do reliably at scale. It can also help us link minor‑effect genes to complex traits. By fusing image‑derived growth curves with whole‑genome data, we can more effectively pinpoint the best parents for the next round of crosses. Incorporating genomics and phenomics, we can predict performance before seed hits the field. Selection models trained on weather, soil, multispectral data, and genotype now forecast which experimental lines will hold up in various environments, potentially saving years of field trials. Together, these tools squeeze the breeding cycle and help us release improved varieties faster. 

UM: Are there challenges in using AI in the kind of work that you do? How are you addressing these challenges?

CC: The biggest hurdles are data quality, generalizability, and compute power. We solve the first by building curated image sets and by partnering with pathologists who verify labels. To keep models from "over‑fitting" one user or camera angle, we train on multi‑location, multi‑year, and multi-populations and run blind validations. High‑performance computing can be expensive, so we rely on USDA’s SCINet cluster and GPU acceleration, which turns a week‑long job into an overnight run. 

UM: Is AI a central part of your work at this point, or do you consider it to still be in an experimental or trial stage?

CC: AI expertise is a valuable tool on the breeder's workbench, but I'm not solely focused on AI research. It’s firmly embedded in our phenotyping pipeline - every breeding season we use AI models to score disease and growth, and those outputs feed directly into genomic‑selection models. Where we’re still experimental is in closing the loop: Allowing the algorithm to autonomously advance or discard lines. For now, a human (breeder) still makes the final call, but AI already drives the shortlist. 

UM: What perceptions or misperceptions about AI do you think it’s important to address, in terms of its impact?

CC: "AI can do a scientists job". There is more to AI than products like large language models (LLMs). While it seems like there's a new development every week, the public is just hearing about marketable products, not necessarily scientific breakthroughs. As a scientist, I work to find the best tool for a job, and that often means building or modifying existing equipment and/or tools to fit the requirement(s). AI has no use without a user or measurable benefit without a purpose. "AI will make mistakes farmers can’t see coming". Like any lab instrument, models are calibrated, validated, and version‑controlled; their error bars are known. "Only big companies can afford AI". Open‑source frameworks and public HPC clusters mean public breeders, like ARS, can deploy cutting‑edge models and release cultivars for public good. 

UM: Overall, how do you think the use of AI could re-shape the science of plant breeding, both for oats and more broadly, in coming years?

CC: With improvements in sensor technologies, I believe AI tools could help monitor and evaluate field trials for traits that are laborious and/or expensive to collect manually. For example, if I have a population planted in multiple environments across the state of North Dakota, that means my team will have to travel and potentially stay overnight at a hotel, multiple times throughout the growing season. Now to make things more efficient, we could develop good prediction models for important field traits using multispectral sensor data collected via drones or satellites. Digital phenotyping would free up time for my team to do other important work like making crosses, analyzing and interpreting these data, and publishing our findings in peer-reviewed journals. 

Down the road, I think there's a chance that whole-season prediction models will be able to integrate weather forecasts with sequence data that will let seed companies and farmers simulate "what‑if" scenarios before seed is even produced and/or acquired. Another benefit of AI is that it can clean noisy data, so growers could soon contribute images from their own fields, speeding regional adaptation, which would contribute to advancing participatory breeding practices. For oats, more efficient data collection means faster progress on long‑standing goals of stable yield, crown rust durability, and improved nutrition.