Skip to main content
A hand holding tweezers picks up seeds on a blue background.

Helping Oregon farmers thrive with smarter seed testing solutions

By Hannah Ashton

Statistician Yanming Di is working to modernize an outdated system for seed purity analysis.

The Willamette Valley is known as the “grass seed capital of the world.” With its ideal climate and soil conditions for growing high-quality grass seed, the region produces more than 90% of the grass seed used in the United States and a significant portion of the global supply.

Being a hub for 500 million pounds of grass seed annually comes with complex challenges, such as outdated testing methods and cumbersome tools — ones that Oregon State University researchers aim to solve. Addressing these problems means farmers would throw less seeds away and have higher quality seed lots.

A multidisciplinary research group is combining expertise in robotics, artificial intelligence, computer science, statistics and crop science to create a modern solution for an outdated system.

“A land grant university is bringing together people with diverse backgrounds and skills to help the people within Oregon. And that is essentially the entire mission of land grant universities,” said OSU Director of Seed Services and collaborator Dan Curry.

For hundreds of years, farmers and scientists have used the same methods to analyze the purity of seed lots. Determined by the amount of weed seeds, unwanted crops and inert materials, seed lot quality impacts every stage of agriculture. To calculate this value, specialized workers use a magnifying glass or microscope to carefully scrutinize a sample. It’s time-consuming, labor-intensive work that invites a degree of human error.

Supported by $255K of grants from the U.S. Department of Agriculture, the Oregon Grass Seed Commissions, and the OSU College of Agricultural Sciences, the group aims to develop a computer vision system for real-time, onsite seed analysis — a tool that could revolutionize farming in Oregon and beyond.

Four people stand in front of a microscope.

Yanming Di (middle, orange shirt) works with the tabletop Ergo Vision to analyze seeds. The researchers take high-quality images of seeds to train the artificial intelligence to differentiate between species.

Eight years ago, members of Oregon State Seed Services envisioned a modern way to inspect seeds. While training an artificial intelligence model to analyze an image is not new, applying this technology to seed purity is. What sounds like a simple task on the surface, actually involves many intricate steps and disciplines.

Before the tool is even developed, understanding the importance of seed testing and the current limitations is crucial, and that’s where Dan Curry stepped in.

When farmers raise a seed lot, they want to ensure customer satisfaction. If weed seeds start growing on someone's newly planted lawn instead of grass, that wouldn’t be good. Or if the seeds aren’t healthy, it directly impacts yield and productivity. Different agencies including the Oregon Department of Agriculture use testing to issue quality tags for seed lots that meet specific quality standards.

When farmers produce a seed lot, they use giant machines to clean out most of the weeds. This requires constant stopping and analyzing the system to make sure they are cutting enough. In other words, throwing away enough to remove the bad seeds. Because growers don’t want to cut too hard and throw away profits, they are constantly grabbing a sample, shutting their machines off and driving miles to a lab.

Analyzing seeds by hand is hard work. It takes three to five years of training to identify up to 200 different seed species and hundreds of hours spent uncomfortably staring at tiny images. Employees who look at hundreds of thousands of seeds each day will make mistakes.

If the grass seed growers of Oregon not only had a more accurate method of testing, but also a portable version, they would throw less away and have higher quality seed lots.

Building on this understanding, a cross-disciplinary research group formed, combining five faculty members, three graduate students and three undergraduates from the College of Science, College of Agricultural Sciences and College of Engineering.

Pictures of seeds use to train AI model.

The artificial intelligence used by the DeepSeed research team learns to differentiate between seed species by analyzing photos like these that only contain one specific seed.

The first challenge is capturing high-quality images of seeds to train the computer to see the differences. Next, it’s figuring out how to maintain consistent conditions while they’re training and testing because if those conditions change, what’s used for training may not apply to testing.

Statisticians like Di are needed to calculate levels of uncertainty, while computer scientists will provide feedback on the neural networks used by AI to perform tasks that typically require human intelligence. Neural networks are algorithms that mimic the human brain’s structure to recognize patterns and make decisions based on data.

In the 21st century, the boundary between statistics and artificial intelligence has started to blur, with both fields analyzing data and trying to make sense of it.

“I don’t really think too much about which area I’m working on, whether it’s AI or statistics. I believe on this team, we just focus on solving the problem,” Di said.

The goal is to have the entire processes automated, requiring the contributions of robotics engineers. To add to the complexity, the group is developing two different versions, the tabletop lab Ergo Vision and a portable light box.

“The idea is we can send the light box to the farmers so they can analyze seeds onsite so they don’t have to send their sample to the seed lab and wait a couple of hours before they can make a decision,” Di said.

The 3D printed prototype currently sitting in the crop science building was made by Ameyassh Nagarajan, an OSU graduate student in computer science and crop science and Logan Snell, an engineering undergraduate.

“I usually work on a lot of theory and engineering, but this is the first time I’ve been involved in something that’s solving a real-world problem,” Nagarajan said.

In the tabletop version, seeds will rest on a stationary flat platform, whereas the lab model incorporates a conveyor belt to transport seeds through the system seamlessly. The tabletop version is designed for high-throughput analysis in lab settings, while the portable light box provides farmers with an on-site solution.

By the start of this year, the group has trained the AI on five types of common seeds. In reality, the system could see a few hundred different seed types, meaning one of the big tasks is to gather more species and introduce them to the model.

Afterward, Di will be involved in working with the computer science collaborators to improve the AI model itself.

“If the machine has say a one percent error rate, it sounds very low. But in practice, the percentage of true weed seeds is also very low. So that means even if you have only one percent of error, that is still a lot of false positives,” Di said.

By applying cutting-edge science to the needs of local stakeholders, Yanming Di and collaborators are turning a centuries-old challenge into an opportunity for multidisciplinary innovation. This collaborative effort underscores the power of science and highlights the commitment of Oregon State to helping Oregonians thrive.