When people think of modern agricultural technology, they often picture driverless tractors, drones surveying fields, or sensors measuring soil moisture in real time. But amid this high-tech progress, one less visible—but equally important—task is undergoing its own quiet revolution: seed analysis.

Today, sorting and classifying seeds is still largely a manual process. Trained experts inspect seeds under lamps and magnifying glasses, assessing their quality—including detecting foreign seeds and physical defects.

“It’s an extremely specialised job. You really need to know your seeds inside and out, and it takes both experience and precision,” says Martin Himmelboe, PhD student at the Department of Agroecology at Aarhus University and the Tystofte Foundation, and lead author of a new scientific paper on the subject.

When the computer joins in

But what if a machine could do the same job—or even do it better? That question forms the foundation of a new publication in Computers and Electronics in Agriculture, where Martin Himmelboe and his co-authors review current research into image-based seed analysis using artificial intelligence.

By applying techniques such as image processing and machine learning, researchers around the world have developed models that visually assess seeds in ways that closely resemble human judgment. Many of the studies show promising results, particularly in classifying weed seeds.

“We’re seeing in the literature that in several cases, the models achieve high accuracy—especially in identifying different cereal and weed species. That suggests the technology could support manual seed analysis,” Martin Himmelboe explains.

The article highlights several studies in which images of seeds are fed into machine learning algorithms that are trained to distinguish, for example, clean from contaminated seed lots. These models learn to tell the difference between pure seeds and impurities—and, crucially, to recognise specific weed species that even experienced analysts might struggle to identify.

A digital future for seeds

It might sound like a niche area within agriculture, but the implications are significant. Denmark is one of the world’s largest exporters of grass seed, and seed quality is essential—for production, export, and certification alike.

“If we can automate seed analysis, we could potentially make the process both faster and more accurate. That would help seed companies respond more quickly, and ultimately lead to more reliable production,” says Martin Himmelboe.

And the potential goes beyond purity analysis. Once the images are captured, new possibilities emerge.

“Earlier studies suggest that in time, we may even be able to identify signs of disease or genetic variation in seeds,” he adds.

People or machines?

Although the technology is promising, it’s not yet ready to replace human experts. Instead, the researchers envision a collaborative approach between people and machines.

“The goal isn’t to remove humans from the process, but to support them. We imagine a tool that seed analysts can use to double-check their findings or to process large volumes of data,” Martin Himmelboe says.

There’s also potential to use the technology as a training tool for future seed analysts. Today, building the required expertise takes years. An automated analysis model could act as a digital mentor, ensuring more consistent training.

It started with curiosity

The idea for the study emerged at the intersection of classical agronomy and modern data science. Martin Himmelboe himself has a background in agriculture and crop production, but has long been fascinated by the potential of technology.

“I find it exciting to take something as hands-on as seed analysis and place it in a modern digital context. It’s not about tech for tech’s sake—but about creating real value for agriculture,” he says.

With his research, he is helping to pave the way for a future where data and digital tools become a natural part of agricultural workflows—even in areas that rarely make the headlines.

What’s next

The new publication is a key step, but there’s still a long way to go before the technology is ready for large-scale deployment. The next research phases will focus on improving model robustness and testing on larger datasets in real-world production environments.

The paper offers a comprehensive overview of where the research stands today—and where it could be heading. Future studies should focus on refining the models, real-life trials, and performance evaluation on larger datasets.

“We wanted to gather and structure the knowledge that’s already out there. Our article shows that the technology holds great promise, and that significant progress has already been made. The next step is turning that knowledge into practice—and it’ll be exciting to see where the field goes from here,” Martin Himmelboe concludes.

More information

Collaborators: Department of Agroecology at Aarhus University
Funding: The project was funded by the Pajbjerg Foundation
Conflicts of interest: None declared
Read the publication: ”Seed identification using machine vision: Machine learning features and model performance” published in Computers and Electronics in Agriculture by Martin Himmelboe, Johannes Ravn Jørgensen, René Gislum and Birte Boelt.
Contact: PhD student Martin Himmelboe, Department of Agroecology, Aarhus University. Email: mhi@agro.au.dk