For the grass seed industry, spotting weed seeds and other unwanted interlopers among thousands of grass seeds is as tedious as it is necessary.

Samples of grass seed are repeatedly analyzed by growers to avoid penalties for weed seeds and other impurities, while seed companies must do the same to comply with labeling rules.

“That takes away time from the rest of their work,” said Dan Curry, director of seed services at Oregon State University. “They don’t want any weeds or any other crops in there.”

By delegating the mind-numbing task to computers, Oregon State University scientists plan to save countless hours of labor without sacrificing seed purity.

Optical tools don’t grow weary as quickly as human eyes, while artificial intelligence software isn’t as prone to wandering thoughts as the human brain.

An interdisciplinary team at OSU is combining those technologies, betting they’ll prove adept at identifying weed seeds and other contaminants in samples of grass seed.

The university’s “Computer Seed I.D.” project also aims to spare seed cleaners and other companies from frequently teaching new employees to discern among seed types.

“The problem is people change within a company or leave the company, and all that training and experience is now leaving,” Curry said.

Examining seed samples is a crucial duty at seed warehouses, which are often owned by farmers who run the crop through filters and other seed cleaning equipment to meet buyer specifications.

If too many contaminants are left in the final product, growers get docked pay, but the removal process risks losing desirable seeds and thus yields and revenues.

Inspecting seed samples helps farmers balance those competing concerns and is a critical job each harvest season.

Unfortunately, the process requires evaluating tens of thousands of grass seeds, which soon tires people out and can lead to mistakes.

“That can be very challenging and tough,” said Yanming Di, an OSU statistician working on the project. “There are particular challenges with grass seed, one of them being that they’re really, really tiny.”

The idea of having the chore taken over by machines came about nearly a decade ago, but the project was officially started in mid-2023 with $255,000 in grants from USDA, OSU’s College of Agriculture, and Oregon crop commissions representing ryegrass, tall fescue and fine fescue producers.

The project team is comprised of 11 professors and students specializing in seed crops, statistics, computer science, precision agriculture, business, biology, robotics and mechanical engineering. They are designing two systems: a smaller portable “light box” that connects to a smartphone app meant for farmers, and a larger device with a higher-resolution camera meant for seed labs.

Using scores of sample images, an artificial intelligence program has been trained to differentiate between ryegrass seeds, tall fescue seeds, and the seeds of curly dock, a common weed.

Eventually, the team expects to calibrate the system to identify each type of seed that it may encounter in samples.

The program’s ability to distinguish between two types of desirable grass seed is promising for the project’s future, said Di. “That’s challenging even for human analysts.”

It’s not yet clear how each system will be sold or otherwise made available to farmers and seed labs. It will probably be another year and a half before the systems are ready to be officially shared with the grass seed industry, as the team is still making improvements.

For images to be recognizable to the artificial intelligence program, the seeds must currently be evenly distributed by hand and the photo background must be removed on a computer, but the team intends to automate these tasks to minimize human intervention.

Another possibility is the development of a conveyor system that mechanically or robotically separates out questionable seeds. That way, human analysts can focus on identifying the weeds or other contaminants rather than hunting for them amid a mass of desirable seeds.

“We still need the seed analysts. It would make their work more efficient,” Curry said.