Gardin is using generative AI and synthetic data to drive plant growth — here's how

• Gardin produces a sensor that monitors plant health and generates growth insights.
• The company initially struggled to source images of plants to train a disease-detection algorithm.
• Generative AI was used to create synthetic images to bridge the data gap and create a model.
• This article is part of "AI in Action," a series exploring how AI is being used across different industries. 

Gardin, an agricultural technology company based in Oxfordshire, England, makes an automated sensor that measures plant health, known as "plant-driven growing."

The sensor collects real-time data about the plants and generates performance alerts and growth insights. With the agriculture industry facing a labor shortage, the technology can help cover the shortfall in workers.

Situation analysis

Gardin's sensors measure plant health with a technique called chlorophyll fluorescence, which monitors how well a plant is photosynthesizing and assesses its level of stress.

While this method can detect whether a plant is healthy or not, it can't precisely identify what is causing the plant stress. The team at Gardin wanted to expand the platform's capabilities so it could classify specific diseases early on.

In order to do this, the team would need to build a machine-learning algorithm. However, Julian Godding, lead data scientist at Gardin, told Business Insider that getting an algorithm to classify plant diseases is "very, very challenging to do." The reason? "There's just so little data," he said.

In order to train a traditional algorithm, Godding said that they would need, for example, 100 images of a particular plant with a disease, and 100 images of the plant species without the disease. While there are lots of images of healthy plants, there weren't enough images of diseased plants to properly train the algorithm, so there was an imbalance in the available data needed.

One solution would be to collect the necessary data — in this case, images of disease plants — manually. However, Godding said that this would be expensive and time-consuming.

"So, you need synthetic data, and that's where generative AI comes in," Godding said.

Key staff and partners

Godding recruited a graduate student at the University of Oxford to work alongside him to build and test a generative AI model. Godding's background is in academia, and he said that looking through published research was his starting point for developing the AI model.

He added that while there was some information already available that they could draw on to help develop their generative AI, they needed to customize it to fit their specific needs.

AI in action

Synthetic data is artificially generated by a computer, rather than collected from the real world. It's previously been used to train models to detect fraud, to resolve the absence of high-quality real-world data relating to fraud.

Some AI experts have said that synthetic data should be used with caution, as it is a "distorted version" of real data. However, consulting firm Gartner estimates that synthetic data will overtake real data in AI models by 2030.

"If you can synthetically create that [needed] data to then train a model on, it saves you a huge amount of time and money," Godding said.

Gardin needed to create artificial images of plants with diseases in order to build a dataset to train their model. They decided to develop this synthetic data in-house. First, the team needed to "prove that it was possible to do this form of adaptation for an algorithm and generalize it in that way."

Godding said that one of the biggest challenges was sourcing a foundational dataset to create the synthetic data.

Did it work, and how do they know?

Over the testing period, they measured the success of the generative AI based on its classification accuracy. It took the team four months to develop.

Godding said that generating the plant dataset with the AI model was "really time consuming and expensive." However, he added, "there was literally no way of doing the alternative."

"The reason that [artificial intelligence hasn't] been realized outside of big tech companies is that it's so expensive to build and maintain AI infrastructure and products that the business case just doesn't stack up in a lot of industries." Godding told BI.

Godding said that once they had generated a good underlying dataset, building the disease detection model was simple. They are now publishing a paper on their work.

What's next?

Looking ahead, Gardin is incorporating artificial intelligence into other aspects of the business. As well as using AI to automate their sensor and turn it into a "mini robot," the data team at Gardin are integrating generative AI into its computer-vision algorithms to ensure that it doesn't overspecify on one type of plant.

This solution means that the model can measure the characteristics of a plant regardless of its surroundings, regardless of "whether the image was taken in a field in Spain, or a greenhouse in the Netherlands," Godding said.