Accelerating seed quality testing with AI
A new deep learning model is helping researchers assess quality faster and more accurately - supporting the delivery of high-quality seeds to growers.
In modern agriculture, seeds are powerful, purposefully designed, and bred to be as productive as possible. Every seed goes through rigorous checks for both quality and vigor (the health and strength of the seed): after all, farmers need to know that the seed they use will work precisely as intended.
Part of the process for testing vigor involves taking seeds and counting the successful germinations - a task usually carried out by technical experts. This has two limitations: it can take time – especially with a lot of assessments to do – and even with expert analysis, natural human subjectivity means measurements can’t be totally accurate.
But what if it could be?
A new patented machine learning process is now making seed vigor assessment much faster and more accurate.
This is the Surface Area BIoMass tool, or SABIMA for short. It captures root and shoot development to quantitatively measure how germinating seedlings are growing.
Developed by Syngenta's Tahl Paran, who works in Global Seed Biology and Testing alongside Pouria Sadeghi-Tehran, a Senior Digital Imaging and Phenotyping Expert, SABIMA represents a step-change in seed assessment.
On the surface, the idea sounds simple, as Tahl explains: “We are using a light box and a camera to create high quality, consistent images of seed germination that have high contrast and low shadow.”
These pictures are then analysed by a computer vision model that measures the root and shoot surface area. These markers serve as reliable indicators of physiological performance.
The original idea came from Tahl’s work in the Seedcare and Biologicals Institute. This is a specialist team where experts test out seed treatments to help protect plants from disease or pests.
Tahl says: “Every so often, we have a product or a recipe (a combination of products) that could reduce quality and so we must mitigate that risk. I was wondering, what if Seed Safety can be done in a more automated way? This is how the idea for how SABIMA came about."
Syngenta's Tahl Paran at work in the lab testing seeds with SABIMA, the deep-learning model he co-developed.
Syngenta's Tahl Paran at work in the lab testing seeds with SABIMA, the deep-learning model he co-developed.
Taking the idea to Pouria, the pair built a deep learning model to measure seedling vigor, and distinguish seeding roots and shoots. With some image annotation and iterative machine learning, the first version of SABIMA was up and running in just eight weeks.
Recently published research has found measuring seed vigor acts as a predictor of real world performance and could help spot the positive effects of seed treatments.
Recently published research has found measuring seed vigor acts as a predictor of real world performance and could help spot the positive effects of seed treatments.
The collaborative duo has recently published research showing that SABIMA measures with the same quality as an expert technician but without any subjective bias. Not only that – their research found that measuring seed vigor can be a solid real-world predictor of seed performance, giving valuable insights faster than ever.
Tahl points out that SABIMA has applications beyond its initial use case. What SABIMA brings is the ability to spot positive effects for biologicals.
“Let’s say we have two seed lots with 100 percent germination, but on one we put an additive that makes them grow bigger. Our model is sensitive to subtle physiological changes that a manual evaluation would miss.”
The standardized set up produces high contrast, consistent images.
The standardized set up produces high contrast, consistent images.
Explaining the impact that SABIMA can make, Pouria says: “This innovation enables more informed and earlier decision making around seed quality and treatment effects because we can measure traits consistently and at scale. It also supports decisions on advancing or stopping treatments, and identifying effects sooner based on quantitative data rather than subjective scoring.
“It’s a real example of digital transformation. Something that is scientifically rigorous, scalable, affordable, and genuinely useful for our research teams.”
