- Potatoes only make up two percent of all arable land in Sweden but account for 25 percent of aggregate pesticide use.
- University researchers partnered with IBM to develop AI-powered drones that can quickly and accurately detect potato blight.
- With accuracy of 75 percent to 97 percent, the drone system can process images ten-times faster than human observers.
Mats Persson, a Business Manager with IBM Sweden, readily admits he did not set out to solve the problem of late blight amongst Sweden’s potato farmers, nor cut down on the nation’s use of pesticides, nor create an entirely new model for crop management.
Like so many innovators before him, Persson, along with a few colleagues at the IBM Client Innovation Center in Malmö, mostly just wanted to play with some high-tech gear.
“We’re all into photography,” Persson told Industrious. “And we basically wanted to figure out if IBM might buy us a drone.”
It was May 2016, and the Cognitive Build challenge, a six-month companywide competition seeking innovative AI projects, was just getting underway. During their afternoon “fika” coffee breaks Persson and his officemates began considering ways to harness drone photography alongside IBM’s latest technology. Persson lives near the Swedish University of Agricultural Sciences, and suggested the team consider farming applications.
One afternoon, Persson ventured over to SLU (its Swedish name is Sveriges Lantbruksuniversitet) and began asking around the cafeteria if anyone was using drones in their work. Multiple people mentioned Erik Alexandersson, an associate professor of plant protection biology, who happened to be enjoying a coffee nearby. The pair struck up a conversation, and Alexandersson explained how his department was utilizing drone photography to detect late blight in potato fields.
Discover how AI and forecasting can protect your crops.
“Just the word ‘blight’ has become synonymous with disaster,” Alexandersson told Industrious, “and for good reason.”
Phytophthora infestans (Greek for “plant-ruining attacker”) is caused by a fungus that rapidly produces spores in wet, mild conditions. Storms can be a potent vector, as are the damp environs synonymous with one of the world’s most infamous infestations of late blight: the Irish potato famines of the 1840s.
Once infected, a plant’s leaves, stems and tubers quickly turn a sickly, mottled black. “Stinking potatoes…” the Irish poet Seamus Heaney once observed, “pits turned pus.” Entire fields can be wiped out in under a week.
A drone’s-eye view of potato fields in Denmark that are infected with blight. Photo by Jesper Cairo Westergaard
In Sweden, such devastation has an environmental impact well beyond crop yields. While potatoes make up less than two percent of the country’s agricultural fields, late blight alone accounts for a quarter of aggregate pesticide use. Swedish farmers have lost more than six billion Euros annually on blighted crops, plus the expense of preemptive spraying and fuel for tractors to spray.
With earlier drone detection, Alexandersson hopes suspect plants can be rooted out or at least more precisely sprayed, sparing their surroundings.
For centuries, farmers could only walk their fields looking for signs of late blight, scanning leaves and twisting stems. As digital photography has improved in resolution, some scientists have experimented with stationary cameras to monitor fields.
Alexandersson and his colleagues realized they could literally take this work to another level. With the more nimble and deployable drones, they can quickly transmit data to a computer or even smartphone. If anything looks suspicious, a human could be dispatched for closer inspection.
While not exactly back-breaking, drone-assisted detection remains arduous.
“It was a lot of imagery for a person to scan through,” Alexandersson said. “And this is just the small number of fields we are testing on. Think of all the farmland in the world.”
Upon hearing this predicament, Persson knew he had found his project.
What if AI could identify the black marks of blight captured by drones, just as it now recognizes traffic lights, faulty welds and discreet cancers?
Spectral photography and color photos are used to enhance blight detection. Photo by Jesper Cairo Westergaard
The SLU already had more than a thousand photos to work with, alongside an unexpected bounty already uploaded into Watson. University researchers in California had similarly been flying drones over infected orange groves. Though studying different diseases, the spectral signatures was similar enough to create a solid foundation, spanning two continents, to enhance the AI’s blight-sighting abilities.
Over the 2017 and 2018 growing seasons, drones were deployed to four fields around Sweden and Denmark. Furthering the project’s educational aims, IBM interns in Malmö led the development work alongside students and researchers from SLU.
In the field, Watson-enabled drones began accurately identifying infected crops three out of four times; in laboratory settings, identifications reached 97 percent.
Some of Alexandersson’ peers still have greater precision—“We call it Watson versus Erland,” he said of his top human blight-spotter—yet they can only review around three images a minute, while Watson can process ten-times as many.
While still years away, Alexandersson and Persson can envision a heartland where drones dart about like birds or insects, autonomously looking for blight and other issues. The most effective bots might even be terrestrial, small rovers programmed to travel along planted furrows inspecting leaves from below, where late blight is most obvious. These devices could theoretically signal larger machines to spray discreet amounts of fungicide or root out infestations.
Drones, data and determined researchers could make blight a rare sight. Photo Jesper Cairo Westergaard
“The economic and environmental implications could be huge for Sweden, and for the rest of the world,” Persson said.
With the fungus utilizing around 25 percent of the country’s pesticides, even a partial reduction would dramatically reduce costs and contamination while boosting yields.
As for the Cognitive Build contest, while Persson’s project did not make the finals—to his chagrin—it ultimately helped IBM win an even bigger prize. In April, Yara, one of the world’s largest producers of fertilizers and agronomy solutions announced a partnership with IBM to build a globe-spanning agricultural platform. Persson’s project was among the demos that closed the deal.
At full deployment, Yara’s platform could service 100 million hectares of farmland, roughly seven percent of the Earth’s arable land.
In the meantime, Alexandersson and his team spent the summer flying drones over test fields outside Give, in Jutland, Denmark, near Legoland. There, a potato breeder had intentionally infected to analyze which spuds were most blight-resistant.
“We picked up so many breeding lines, in such concentrations, it’s going to be amazing for the data,” Alexandersson said, excited by blight as only a researcher could be.