Lightning-fast AI tray assistant at Bogrundet

Each year, the sowing operators at SCA's plant unit NorrPlant handle approximately two million seedling trays to sow the roughly 100 million seedlings that SCA contributes annually to the forests of the future. Until now, the detection of damaged trays relied on manual inspection, or action was taken only after a production stop had already occurred.

“Identifying and removing damaged trays is an important task. But it has also been somewhat of a work environment challenge, as it is demanding to stand and inspect and manually remove trays all day,” Anders explains.

Bogrundet’s in-house developed AI model is now fully operational, following extensive work by Matthias Larsson, Business Development Specialist within Technology and Digitalisation.

“We saw clear potential in an AI model. It was also particularly timely since the production line was being upgraded ahead of the 2026 sowing season. It has truly been both enjoyable and rewarding,” says Matthias.

Trained on 8,000 images

Matthias invested significant time in defining how the AI model should function before building the interface and testing different camera solutions. He then spent many hours manually training the model to recognise damage such as broken flaps and chipped outer edges, enabling it to identify defective trays accurately.

“Collaboration with the operators has been invaluable, as we needed a shared understanding of what constitutes damage and what tolerance level should apply. Initially, the model reacted to dirt and debris stuck on the trays, but it no longer does. I’ve probably used around 8,000 images to show and train the model to recognise which types of damaged trays should be sorted out,” Matthias explains.

One tray per second

A crucial aspect of the development process was ensuring the model could detect trays at high speed.

“A large number of trays are used, so time efficiency was essential. Our goal was for the model to process one tray per second — and we achieved that,” Matthias emphasises.

The AI model scans all trays at the beginning of the production line. As each tray passes along the conveyor belt, it is photographed using the system’s photo cells and cameras. The AI then determines whether there is damage, such as a broken flap or a chipped edge.

“When a damaged tray is detected, it is diverted to a separate track and sorted out, while intact trays continue along the production line,” Anders explains.

He is very pleased with the new assistant and commends the work behind it.

“I expected we might encounter a number of initial challenges, but it has gone far better than anticipated. As defective trays are removed, we improve the overall quality of our trays, which is very positive,” Anders concludes.

Top picture: Michael Engman.