Blending advanced AI and ‘old school’ tools to assess Fusarium head blight severity

Like virtually every industry today, agriculture is going big into artificial intelligence (AI), exploring ways AI can make farming more efficient and profitable. Yet, while AI has its benefits, University of Guelph professor and ASA member John Sulik thinks sometimes researchers apply “AI for AI’s sake” when simpler tactics would do. 

Case in point is the approach he recently developed with his University of Guelph colleagues, including the project’s leader, master’s student Riley McConachie, and wheat breeder Helen Booker, a CSSA and ASA member. Reported on last year in The Plant Phenome Journal, their technology combines an advanced AI technique—deep learning—with a traditional tool—the spectral index—to quickly and reliably assess the severity of the fungal disease, Fusarium head blight (FHB), on wheat heads. 

The team could have gone for an end-to-end deep-learning approach, Sulik says, which “would have been more impressive to deep-learning enthusiasts.” But that wasn’t the goal. “We just wanted to make a model we could use to improve wheat-breeding efforts and, for growers, to rate the severity of FHB in their fields.”

Going beyond visual assessments for more objective scoring

Today, plant breeders and farmers assess fungal diseases like stripe rust and FHB visually; in other words, they eyeball the plants and then estimate the percentage of infection. This is straightforward for stripe rust, where plants are either obviously infected or show no disease at all. But FHB in wheat is another story, Booker says. 

The point of isolating these pixels is to analyze them further—for disease, for example. And for this last step, Sulik proposed using an “old school” tool, the spectral index. To rate FHB severity, the technique uses specific red and green bands of light, captured by a cell phone camera, to calculate the degree to which each wheat head pixel is bleached—signifying infection—or green—signifying health. Each pixel value is then compared to a threshold value for disease experimentally determined by McConachie. Pixels falling above the threshold are scored as healthy and those below it as infected. Finally, disease severity is estimated by counting the number of pixels above the threshold and then dividing by the total number of pixels in the wheat head image. 

In summary, Sulik says, “David's approach got us to the heads. Then we used the index to see how green each pixel was and to see if there was a boundary, a pixel value, we could use repeatedly” to score disease. McConachie went on to show the method is in fact highly repeatable. He has now bundled everything into an app that plant breeders and farmers can eventually use in experiments and on farms.

Sometimes AI isn’t needed

Through the project, McConachie also became something of an expert in AI techniques. He learned how to work in the programming language, Python, for example, and with open-source deep-learning models, such as Meta’s “Segment Anything” and models available through Global Wheat Head Detection competitions, led by David. But the biggest lesson he learned is that sometimes AI isn’t needed. 

“Several past studies have tried to use complete deep-learning workflows to estimate FHB severity, but they didn’t transfer well to other datasets,” McConachie says. “By using deep learning as a stepping stone to get the data—in this case, images—to the point where other techniques could be used—spectral indices—we created a much more repeatable and transferable solution.”

Sulik agrees, adding that making the project “more AI” would have also made it more time-consuming and expensive. Instead, the team focused on “… getting a solution that's good enough at the lowest cost,” he says. “That's what agriculture is about.”