BENTON HARBOR, Mich. — Lynn Sosnoskie, associate professor in the school of integrative plant science, horticulture section with Cornell AgriTech, says growers often call her “Dr. Doom” because she says, “the weeds always win.”
During the Great Lakes Tek Flex, Sosnoskie took a deep dive into the potential of artificial intelligence and technology for the future of weed control in specialty crops.
The Limits of Conventional Herbicides
While she says herbicides are the most widely used weed management solution, they’re often the go-to solution for growers due low cost, fast action, and simple application.
“We put out more herbicides to more acres on more farms than insecticides and fungicides combined,” she says.
But, she says, the current model of use is under threat from long runways for the product pipeline, driven by high development costs and extended release timelines.
“It can take more than a decade and a hundred million dollars plus to bring a new pesticide from discovery to commercialization,” she says. “We don’t have the same number of releases that we used to have. So, we just have fewer products available.”
And, growers wishing to export face tough maximum residue limits (MRLs). There are also environmental considerations, potential crop injury, worker exposure, and more that can also impact herbicide use. She also says there’s changing social and environmental pressures around herbicide use and technology choices.
“It doesn’t mean that herbicides are necessarily going to disappear,” she says. “But it’s going to mean that I think the market for alternatives is expanding again.”
There’s also strong concern about future herbicide resistance, which she says is going to continue to put strain on the existing chemistries.
Beyond Chemical Control
All this, she says, is fueling a focus on new technologies for weed management. These solutions combine cameras, computing, and engineering to mechanically remove weeds, target sprays, or use lasers and energy sources to eliminate unwanted vegetation.
“When we’re talking about precision, we’re talking applying the right tool in the right place at the right time,” she says.
Sosnoskie says the value of AI-based precision weed technology is not only effective weed control but also improving crop vigor and yields, reducing labor and herbicide use. She says a grower she met at FIRA a few years ago told her his only motivation for adding precision weed control was to keep costs from escalating. In fact, he didn’t think the precision weeding would reduce costs, but told her, “’Realistically, I just want to try to prevent increases that I can’t keep up with in my costs.”
One thing she points out that as growers integrate precision weed control into operations, there’s often additional challenges that arise. An example of this is a study in a commercial onion field where researchers converted the field from a high herbicide program to substituting laser weeding for many of the applications. This helped improve onion vigor. While a more vigorous plant may on paper seem to be more resistant to onion thrips and stemphylium leaf blight, the planting that reduced herbicide use had more disease pressure because the canopy closed in faster than anticipated.
There’s additional metrics growers should also look to when integrating precision weeding such as water quality, soil health, biodiversity, habitat protection and worker safety. This includes workers being exposed less to the herbicides but also to heat, UVs, dust and stress.
Using precision weeding might also mean growers can have more diversified crop rotations as some herbicides have residual effects that can carry over to future seasons.
Growers might also gain access to new markets thanks to meeting export MRLs.
Geographic and Geological Hurdles
Sosnoskie notes that adoption challenges often depend on farm location, with larger operations adopting technology much faster than smaller ones.
“Those larger farms are not evenly distributed across the United States because in the Eastern U.S. our medium farm size is 150 to 800 acres, but the large farms really tend to be in the western United States which are 1,500 to 5,000 plus acres,” she says.
Technology developed in the West may not easily translate to Eastern or Midwestern farms, she says, based not only on geology but also European settlement patterns. She highlighted snapshots of two very distinct farming communities: Lancaster County, Pennsylvania and Yuma, Ariz. The farms in Yuma are much more consistent and rectangular, whereas farms in Lancaster County are far more irregular.
“Those are very different environments and it affects how well technology scales and the efficiency of the technology,” she says.
Soil on Eastern farms are much rockier which impacts operational efficiency. She says she worked with one ag tech company that had smaller robots, which are ideal for Eastern farms. The issue was the vibration caused from driving over rocky soil voided the targeting accuracy of the technology.
Smaller, scattered fields also mean increased transport time between plots, she says.
“That means there’s higher nonproductive operating time,” she says. “There’s greater overlap potential and that’s inefficiency and there’s increased route planning complexity and that could make autonomously swarm operations fleets are going to be harder to formulate.”
Also, these rough terrains mean increased wear on sensors and implements, issues with navigation guidance, LiDAR nterpretation, background vegetation, cameras, etc.
“What you build and what you design and what you optimize in California is not necessarily going to do you a service here,” she said to the equipment manufacturers in attendance.
She says the future of agricultural AI requires equipment developed to perform reliably across diverse environments.
“Making it more adaptable to boundaries, obstacles, slopes, and changing field conditions and maybe self-monitoring,” she says. “Recognizing declining accuracy, sensor wear, and unsafe conditions and thinking about the economic scalability of small and medium-sized growers.”
Biological Weed Adaptation
And, she says, harkening back to her “Dr. Doom” nickname, there’s always a tradeoff and “weeds always win.” She says as technology seeks to replicate human decision-making and human weed removal, weeds over time become more difficult to distinguish from the crop itself.
She points to a study in rice that identified Vavilovian mimicry, where weeds evolve to resemble co-located crop plants through unintentional human selection. Over time, humans have removed the weeds that are easy to select and left the weeds that are harder to identify. She worries that vision systems might impose the same selection pressure on weeds in the future. Or, weeds could evolve to emerge after the crop and use the crop’s structure to hide it from vision systems.
“For every action there’s an equal reaction and with our AI systems, there’s going to be a reaction to our use of technology,” she says.
Sosnoskie also pointed to a study conducted by Jason Norsworthy at the University of Arkansas that discovered that reduced or sub-lethal herbicide application often leaves surviving weeds exposed to lower doses. This creates conditions where surviving weeds are more likely to develop resistance.
“Just because we’re using a new technology to put our herbicide application out doesn’t mean we’re changing the selective pressures that facilitate resistance development and I think it’s stuff that we need to be mindful of going forward,” she says.
She says she is part of a team developing a weed management algorithm from scratch and has focused intensely on the crop the algorithm will be used in as weeds often present differently under different stresses and injuries.
“We’re focused on the easy one, which is the crop, because we failed when we tried to do it on the weeds,” she says. “Our weeds were all over the place with their appearance and their shape.”
Economic and Regulatory Reality Check
Other things to think about are the slow pace of the regulatory environment. She rhetorically asked whether each herbicide would need a specific label and use rate for each different piece of weeding equipment.
She also highlights a central tension facing agricultural tech companies: balancing investor expectations with the needs of the growers they serve. Because scaling these technologies requires significant capital, she says the industry should be concerned about whether the current venture capital environment is patient enough to support long-term development over short-term returns.
When asked what the industry needs most to accelerate technology and adoption, Sosnoskie says, “We need more proving grounds.”
“We need the support to actually build the sites. that the growers need, that the industry needs to actually be able to put everything through its paces to do it effectively through as many places as we can to get the data out as quickly as possible.”
