A few years back, a team of experts – in computerization, engineering and agronomy – got together in Silicon Valley to brainstorm about weeds and how to kill them more effectively. Something new and exciting eventually followed.
They developed a smart sprayer that applies herbicide only on weeds it detects, leaving the crop and bare ground unsprayed. That, in turn, can significantly reduce herbicide use, depending on weed density. Beyond that, this approach could expand the modes of action available to farmers, since the system only sprays the weeds.
Think of it as a sort of site-specific, post-directed sprayer.
The proof of concept – Blue River Technology’s (BRT) See and Spray unit – mounts on a pull-behind tool bar, and it will continue to be tested in U.S. cotton and soybean fields in 2018. After this coming summer’s evaluations, the company expects to have a better idea of when the technology might be available to farmers.
In ag engineering circles, folks are taking the concept seriously – enough so that Deere and Company bought BRT last fall for a hefty $305 million. Deere plans to keep the 60-person team in its present location in Sunnyvale, California, as the development process continues.
The system includes three major components – a simple digital camera, a computer to analyze and trigger spraying and a set of spray nozzles specifically designed for the technology.
Teaching The Machine Like You Would A Child
Once the basic design was completed, BRT technicians began to “teach” the sprayer the difference between weeds and crops through a process called machine learning.
“We take hundreds of thousands of sample images of different fields and different weeds, like pigweed and horseweed (marestail),” says Willy Pell, director of new technology at BRT. “We bring those images back to a team of humans who draw boxes around weeds and crops and label them. With that information, we train our model to tell the difference between crop and weed.”
To improve accuracy for real-farm situations, sample images include crops that were drought-stressed and hail-damaged.
“Over time, you’re exposing the model to pictures it’s never seen before and it can still detect the difference between crop and weed,” Pell says. “It’s much like you’re training a child to do well on a test. You give him lots and lots of practice problems, and the problems he faces on the test are roughly similar to the practice problems.”
Targeting Weeds One Frame At A Time
As the sprayer moves through the crop, the camera constantly snaps digital photographs of the ground. The overlapping images are approximately 40-inches wide across the row and 15-inches deep in the direction of travel. Image recognition software, similar to facial recognition software used in Facebook, analyzes the shapes of leaves inside the image and determines whether they are friend or foe.
If the computer determines that a weed is present, it activates the sprayer. The nozzles lag the camera lens by a couple of feet. A second camera behind the spray nozzles verifies the effectiveness of the application. With the help of the rear camera, the computer can self-calibrate to correct for any errors.
“The entire process is completed in 100 milliseconds, which is faster than you can blink your eye,” says Pell.
Learning Never Stops – New Day, New Lessons
While the technology is a form of artificial intelligence, “it’s more like really good memorization,” Pell says. “If you start getting too far out of the bounds of what it has memorized, it’s not going to work as well. The trick is figuring out what it has to memorize.”
In lab tests in it’s current configuration, the sprayer has about 99% accuracy for detecting weeds in cotton and wheat, Pell specifies.
The technology continues to learn about the shapes it sees in the field, he adds.
“If we’re not getting optimum performance, we can examine those images with human eyes and retrain for the next day. The system is designed to improve over time. It loves stress, because it performs better the next day,” Pell says.
In situations where a weed and the crop are entangled, “a grower can set a risk threshold based on whether he values controlling escapes more than sustaining a little crop damage or if he really doesn’t want to damage the crop at all. That said, we are spraying the weed very precisely, within a half an inch,” Pell explains.
Eventually, the machine can be taught to handle other types of applications. “You can flip the technology to apply fungicide or fertilizer on the crop,” according to Pell. “Ultimately, we want to optimize every plant – treat every plant with exactly what it needs, no more, no less.”
Potential Benefits To Producers
Pell says producers can significantly reduce herbicide use with the system. “We’ve seen a 90% reduction in herbicides in most fields, although your herbicide reduction will totally depend on weed pressure.”
With the ability to apply labeled, broad-spectrum herbicides with the machine, producers could increase their arsenal of modes of action, Pell points out.
“It might allow the use of herbicides that have never been released simply because they are too expensive to produce,” he continues. “But if you can get a 10X reduction in herbicide use, that exotic material could be back on the table.”
Pell notes that one major ag chem company was an early investor in BRT and his group is working with other herbicide manufacturers to ensure that See and Spray units works with a variety of products already on the market.
John May, Deere’s president of Agricultural Solutions and chief information officer, says the investment in BRT is similar to Deere’s acquisition of NavCom Technology 19 years ago. That gave Deere an immediate and early presence in the agricultural GPS market.
Editor’s Note: Trimble manufactures the WeedSeeker Spot Spray System. Although it cannot tell the difference between weeds and a crop, it does use light reflection technology to identify weeds and trigger spraying. Here’s a link to further information.