Artificial intelligence for mussels
Imagine being able to photograph a freshly picked iconic Greenshell mussel off a mussel line and know instantly whether it is in peak condition and ready for harvesting for food or even better, harvesting for its bioactive content.
The innovation team at Sanford have been working on artificial intelligence technology that can do just that. It has the potential to help marine farmers decide when to harvest and what the best outcome is for the harvested crop. This new technology also has potential for use in other parts of the primary sector.
The technology is needed because even the best trained human eye isn’t able to assess all the variables that can determine how good the mussel is for nutraceutical use.
Sanford’s General Manager of Innovation, Andrew Stanley says, “our mussel sourcing teams and third-party suppliers are highly experienced, but what we are looking for is really difficult for any human to measure, particularly while out on the water. Providing new high-tech tools to those making harvesting and sourcing decisions is going to help improve quality and consistency”.
The approach involved gaining a better understanding of how individual mussels vary. The team selected mussels from a range of areas across the Marlborough Sounds and elsewhere and measured multiple features over an extended period. The team spent many hours dissecting and testing mussels, including using Near Infrared Spectroscopy (NIR).
Logan Nutsford joined the Sanford innovation team in December 2019, bringing with him new skills and knowledge in automation, mechatronics, and artificial intelligence. Logan set about delivering a working prototype of a system which would prove artificial intelligence could predict outcomes for mussels. He was able to show, after several weeks of work that AI could in fact learn to understand quality.
The next step was building the first version of a photographic and sensor system which would deliver repeatable and consistent results. Because neural networks can be trained more accurately by removing changing backgrounds, Logan designed a vision box that controlled the lighting and other conditions for assessing the product. These consistent images had the added benefit of allowing time-lapse capability and size assessment. Building the system involved using a combination of custom designed 3D printed components and off the shelf parts.
Over several weeks, around 1000 mussels were photographed and tested. The final output of these trials was 90% accuracy in predicting the quality outcome.
The Sanford team have progressed to the next version of “mussel vision” which now features two cameras, a weigh scale and a touch screen with an improved user interface. The on-board computer is capable of capturing sample location and other details, measuring length, width, height and weight, and providing an assessment of quality using model predictions.
The platform has been designed to be duplicated with relative ease to allow multiple machines to gather data about crops simultaneously, with units linked to a central database. The future goal is to make these devices so that they can be used anywhere, in any condition for a range of purposes.
The Sanford-developed technology may have application in other industries according to Andrew.
“In the primary industry often individual plants and animals are subtlety unique. The differences may not be visible to the untrained eye, but that doesn’t mean they aren’t significant. This technology can identify those differences and assess how we can best take advantage of them.
“We are yet to fully explore how far we can go, but who would have thought, even a year ago that AI and Greenshell mussels would be spoken about in the same sentence.”
