Engineers are harnessing artificial intelligence to plow the fertile new field of precision farming.
By Thomas K. Grose
Agriculture has always been a precarious undertaking. Scourges from beetles to erosion to war can doom harvests and spawn famine, and even today’s industrialized farms remain vulnerable to the whims of weather. Weeds, which compete with crops for water, light, and nutrients, are a perennial foe. Hundreds of species already have become resistant to herbicides, and a 2018 North Carolina State University study determined that evolving defenses will eventually render chemical weed killers useless.
Girish Chowdhary, an associate professor of agricultural and biological engineering and computer science at the University of Illinois Urbana–Champaign, offers an alternative to this herbicidal arms race: deploy platoons of small robotic weeders guided by artificial intelligence. Weeds cannot evolve defenses against mechanical plucking, observes the director of UIUC’s Illinois Autonomous Farm. Moreover, the robots—which can service more than 2,000 acres—promise to greatly reduce the environmental harm caused by the overuse of herbicides. “This is something everyone wants,” Chowdhary says about the potential to end the need for herbicides once and for all. While his device is still in the early stages of development, he stresses that field tests and models “show it should work—it’s totally possible.”
Robo-weeders would join a growing phalanx of AI-enabled innovations plowing experimental new ground in the rapidly advancing practice of precision agriculture. Born in the 1990s with John Deere’s introduction of GPS-guided tractors, digital farming has blossomed to encompass a host of computer- and sensor-enhanced tools being developed by academic, government, and industry researchers, including engineers and plant and soil scientists. Their shared goal: turbocharge the next green revolution via pinpoint delivery of water, fertilizer, and other precious resources, thereby ensuring the world can feed its burgeoning population—projected to hit 11 billion by the end of the century from around 7.9 billion today. The federal government is making heavy investments in developing these technologies, but US researchers believe they will also have global applications.
“It’s about improving the health and yields of crops,” explains Cherie Kagan, a professor of electrical and systems engineering at the University of Pennsylvania who is leading a recently established National Science Foundation Engineering Research Center for the Internet of Things for Precision Agriculture (IoT4Ag). “The more crops we can squeeze out of each drop of water and energy the better,” she adds. As Chowdhary puts it: “We do not need to grow much more but do it smarter.”
Seed Money
Beyond boosting harvests, AI-assisted agriculture promises to pay broad ecological and economic dividends. Potential environmental benefits range from decreased use of water, pesticides, energy, and nitrogen fertilizers to the breeding of resilient hybrids and other climate-related interventions. A “game changer” for food security, infrastructure, and the environment is how Kei Koizumi, chief of staff of the White House Office of Science and Technology Policy, described precision agriculture at ASEE’s annual Public Policy Colloquium in February.
The federal government’s investments include $26 million in Kagan’s multi-university initiative. It’s one of five agricultural AI projects that NSF has jointly funded with the US Department of Agriculture over the past two years. Industry also is betting the farm. A Google search for “ag AI start-ups” turns up scores of new companies worldwide. Market research companies predict that agriculture AI is set to skyrocket. For example, New York-based Facts & Factors estimates that the market for agriculture AI will hit nearly $4 billion by 2026, up from an estimated $1 billion in 2020.
Still largely at the blue-sky stage, ag AI builds on decades of research in disciplines as diverse as robotics, plant genetics, and climate science. However, growers already have reaped enormous efficiencies from smart technologies. GPS-equipped tractors, for example, steer automatically along fields, minimizing fuel use and misplaced seeds. The combination of soil moisture monitoring with weather forecasts has increased the efficiency of irrigation systems. And a Microsoft app can guide farmers on optimal planting dates via text message.
Still, the sector remains ripe for radical change. The World Wildlife Fund, a nonprofit environmental group, estimates that 15 percent of all food never gets harvested or makes it off the farm due to causes such as labor shortages, weather, and cosmetic imperfections. That equates to about 15 million tons of produce annually—nearly a third of total US food loss. And excessive use of chemicals and fertilizer contributes to groundwater pollution. According to the Environmental Protection Agency’s National Water-Quality Assessment, agricultural runoff—which includes some of the 500,000 tons of pesticides and 12 million tons of nitrogen fertilizer that US farms apply annually—is the leading source of impairment to rivers and streams and second-largest to wetlands.
Data Harvesting
Just as advances in digital technologies and automation are driving a fourth industrial revolution, big data and computer modeling promise to transform food production from seed to table. Engineering academics seek to accelerate the transformation. They are experimenting with small autonomous vehicles that can patrol rows of soybeans for incipient weeds and testing plant “tattoo sensors” that detect hydration in corn leaves. Ultimately, researchers aim to create cheap, easily installed smart technologies and computer models that will provide farmers with maps and other guidance on which areas need water, fertilizer, or pesticides “so they don’t just water or spray everything,” explains Kagan.
Agriculture 4.0 faces unique growing pains, however. For instance, fields are more challenging environments than offices or factories. Consider sensors, a key component in AI-assisted farming. Unlike industrial versions, instruments that measure pH, iron, and nitrogen levels in soil or interpret images from cameras deployed to visually inspect plants for infestations aren’t cheap. That’s because they must operate in large, often harsh outdoor environments, says Kagan. A single sensor that can measure soil moisture and chemical levels in roughly one hundred acres can top $100, compared with an average of 38 cents for industrial IoT sensors. Moreover, smart ag sensors are comparatively hefty, and some need electric cables to power them, “so costs can add up quickly,” observes Kagan.
Smart Chaff
IoT4Ag’s researchers seek to reduce such barriers by designing tiny biodegradable sensors that cost pennies apiece, require little or no power, and can be scattered around plants like seeds. The “smart chaff” would communicate state-of-the-field information via optical or radio-frequency signal. Nontoxic materials will allow the sensors to provide data on root growth or the hardiness of new hybrid seedlings, yet still be affordable enough to permit seasonal renewal. Once the data are stored in the cloud, algorithms would be deployed to build models that incorporate weather forecasts and other relevant information to produce maps and other graphics that farmers can consult as easily as a smartphone app.
Challenges abound. For starters, the quality of the data varies greatly and arrives at different times from many locations. “How do you bring together all these signals, how do you fuse them together?” asks Kagan. Investigators also must figure out how to retrieve and process data from specially designed root-monitoring sensors. One option, Kagan says, is using passive optical or radio-frequency sensors that require little or no power to operate and change color in response to a chemical target. They can reflect that signal change back to sensors aboard aerial or land-based robots or other farm equipment.
Conveying data from the farm to the cloud, where they can be analyzed, fed into models, and used to inform farmers on irrigation, pest-control, or transplanting needs, poses another crucial challenge. “Most fields are in the middle of nowhere, where you have limited communications,” notes Kagan. Her team is examining ways to relay data “in hops” from one robot or farm vehicle to another using local wireless technologies, then uploading to the cloud from a broadband-equipped farm building.
Digital Doppelgangers
The AI Institute for Resilient Agriculture (AIIRA), a $20 million NSF/USDA multi-university initiative based at Iowa State, seeks to customize crop care through the creation of exact computerized copies of individual plants or entire fields. For crop management, ground robots, drones, and satellites would glean a wide variety of real-time data—including microclimates, field slope, and soil hydration and nitrogen levels—from sensors and cameras to create digital twins of whole fields. Plant breeders would use sensor-laden robotic arms to collect data from individual plants that then would be fed into each digital doppelganger. The “aspirational goal” is predictive models, says George Kantor, a robotics research professor at Carnegie Mellon University and one of AIIRA’s directors.
The models would provide farmers with a “decision-support tool to help them decide which varieties to plant in which parts of the field,” explains Kantor. The result: increased yields while reducing the negative environmental impact of water and chemical use. The model would include economic data on likely market prices, permitting farmers to “pick varieties that not only yield the most but have the highest market value so they can maximize profits,” says Kantor. At the individual plant level, the models would give breeders new and more detailed information about plant genetics and physiology to assist in the development of hybrids able to withstand warming temperatures, droughts, and pests. The models would hypothesize how the traits of a specific plant enable it to fare when exposed to hypothetical environmental changes. Breeders already have access to decades of models of individual plant parts and processes, such as photosynthesis, Kantor explains. AIIRA researchers will use machine learning to try to unify all these models into one more comprehensive model, starting with corn. The models will be “agnostic,” says Kantor, and could be applied across a variety of crops.
Disease and Drought
At the University of Illinois, a five-year, $20 million NSF/USDA-funded research institute is focusing heavily on computer vision and machine learning. Investigators at Artificial Intelligence for Future Agricultural Resilience, Management, and Sustainability (AIFARMS) are developing camera-equipped drones and ground-based robots that can maneuver beneath crop canopies and detect disease and infestation as well as measure such physical properties as leaf width and height, or the size of corn ears.
The institute is also working to address the effects of drought, with approaches tailored to particular regions. As AIFARMS head Vikram Adve notes, rising temperatures driven by climate change can increase rain in some areas but cause drought in others. The computer science professor is using AI to improve the study of plants’ basic biology, in order to determine why some species are better able to tolerate drought and then develop hybrids that can mimic that trait.
UIUC’s Chowdhary is also involved in this work through a $7 million NSF/USDA project at Iowa State University focused on cyber-physical systems, called COntext Aware LEarning for Sustainable CybEr-Agricultural Systems (COALESCE). Test fields, he notes, are filled with a wide variety of hybrids. Using robots to measure and assess their physical properties “is a lot faster and cheaper than doing it with humans.”
COALESCE’s team hopes to use robotics, sensing, and modeling to ratchet up advances in a field it calls “ultraprecision agriculture,” defined as working at the individual plant level to find and eradicate problems in microplots early on. Chowdhary aims to reduce herbicide use first by 30 percent, then 70 percent, and ultimately up to 100 percent, and also to decrease the use of fertilizers by about a third.
Learning Machines
Another area ripe for labor savings and greater efficiency is farm equipment. Large machinery already is “heavily instrumented,” says Adve. Still, there’s much room for improvement. Although today’s tractors are largely self-driving, humans must stay aboard to monitor the screen and avoid large rocks or other impediments. “Computer vision can increase the level of autonomy” and “free up humans for more productive work,” he says. Similarly, automated weeders exist today, Adve explains, but their use is limited “because farmers have to walk behind them, so they’re labor-intensive.”
Building computer-vision robots that can sow cover crops—plants grown to keep fields healthy between seasons—is an initial step toward reducing the need for human minders, Chowdhary explains. Studies show that cover crops can reduce the need for fertilizers, increase yields, and naturally sequester carbon dioxide, yet only a tiny percentage of farmers use them because of the effort and costs involved. “Doing it robotically makes it cheaper and hassle-free,” he says. Cover crop-planting robots that can roll under the canopy also could encourage a cultural shift to planting a variety of crops on the same land. “Monoculture worked for centuries but is now seen as environmentally wasteful and making climate and water problems worse,” underscores Chowdhary.
The challenge that Chowdhary’s team and his ag robotics start-up EarthSense are tackling is demonstrating the ability of seed-planting robots to successfully traverse hundreds of acres of farmland. That proof of concept should lead to commercial planters and then to weeders, he says. In addition, one of his PhD students is working to develop robots that can vacuum bugs from plants. “It’s really all attachments, once we have nailed a platform that can navigate huge fields,” he emphasizes.
AIFARMS is the only federally funded engineering research institute with a livestock component. Cattle and pig farms typically have thousands of animals to keep track of and only a handful of people to do it. “Sheer scale makes it difficult for humans to keep up with animals on an individual level,” says Adve. Stress- and sickness-inducers usually can be remedied if detected early enough, so he and his team are working on computer-vision systems that autonomously detect animal behavior and use AI to draw up individual profiles and predict outcomes. For instance, a pig that lies down too much is likely suffering from a cold or infection. Distinguishing or keeping track of one ailing creature in a group or herd is a tall order, however, especially when the animals huddle close to each other. Adding to the problem, Adve explains, “pigs often don’t have individual markings. Cows are easier.” However, he continues, cameras can use and combine other physical attributes to track pigs, including patches around their eyes and body color variations, size, and shape. Video also could help identify animals by gait. “The bottom line,” Adve acknowledges, “is that [livestock tracking] is difficult enough that it is still an open research problem.”
Berry Good Harvesters
While corn, soybeans, and other row crops have garnered the bulk of attention, NSF/USDA’s AI Institute for Transforming Workforce & Decision Support (AgAID) is zeroing in on fruits, nuts, and berries—specialty perennials that usually require irrigation. Led by Ananth Kalyanaraman, a computer science professor at Washington State University, the $20 million, five-year initiative aims to apply the increasing amounts of data generated as more orchards and vineyards deploy sensors to help farmers better manage crops throughout the year. Improved or new models derived from the sensors could indicate ideal times to prune and harvest, for example.
Sensor data will include soil conditions, pH, and weather forecasts, including wind direction and speed. Visual data will depict leaf and flower color and fruit quality. Because there is “concrete evidence of tremendous variables, even on the same farm,” Kalyanaraman aims to “adopt solutions that meet individual farms’ needs.” He envisions each farm employing a data manager who helps the grower decide where, when, and what resources to deploy. Higher yields aren’t always the main goal. Because of worker shortages, many fruits go unpicked, so “there’s a lot of wastage,” says Kalyanaraman. Novel models would enable farmers to better match expected yields to anticipated labor needs and improve quality.
AgAID also is developing regional models to better predict water availability and help state and local planners decide how much water they can safely allocate to farmers. That’s important, Kalyanaraman says, since the farmers “are looking for early-season information on water availability.” Another AIFARMS project involves using computer vision to guide robotic arms that can pick fruits and berries growing close to the ground without crushing them.
Love Thy Laborers
Won’t ag AI and automation put farmhands out of work? “That’s a common misconception,” contends Kalyanaraman. Aditya Johri, a professor of information sciences and technology at George Mason University who is involved in both the COALESCE and AIIRA projects as an education outreach expert, points to a longstanding trend of attrition, at least for row crops. “Most farming in the United States is already extremely machine-intensive and automated,” he notes. Moreover, he adds, many young people are abandoning farm communities for greater opportunities in cities, hobbling growers of labor-intensive specialty crops and vegetables.
“Humans will still be an integral part of things,” maintains Kalyanaraman. “We’ll be empowering the existing workforce with AI.” Because farmhands will be needed on the automated fields of the future, education and training is a core component of all five NSF/USDA projects. AIFARMS, for example, is developing online master’s degrees aimed at working professionals and next year will launch online options for individual certificates. The programs will share courses such as Mobile Robots; Autonomous Decision Making; Crop Growth and Development; and Agriculture, Food, and the Environment. Additionally, both IoT4Ag and COALESCE are partnering with USDA extension services and nonprofit organizations to develop ideas for how best to train the current workforce to use AI technologies.
US researchers say their labors will also bear fruit in global applications and are confident that their innovations can be cost-effective for developing countries. Kantor explains that the digital twin project has “overseas applications in mind,” and researchers are working with non-funder overseas investigators. Technologies will enable breeders to tailor crops to specific climates worldwide—for example, making them more resilient to warmer temperatures.
Johri, a Prism columnist, notes that there’s little risk of the innovations putting overseas or domestic farm laborers out of work. Like their US counterparts, young people in developing countries like India are migrating to urban centers for job opportunities, he says. Johri also is convinced that as the technologies are adopted, their costs will fall quickly—particularly because academic researchers in developing countries are also busy commercializing ag AI inventions. “Around the world,” he says, “there are a lot of start-ups, hundreds in India alone.”
Whose Data?
Coaxing a significant number of farmers to take up these nifty new AI tools will, however, require engineers to sharpen their marketing skills. Adve acknowledges that while some farmers are pioneers, others are “wait-and-see” and some even “not-in-my-lifetime.”
AI researchers also admit that the pace of acceptance could be slowed by one looming issue: the collection of massive amounts of data. Accordingly, data privacy and ownership are becoming hot topics, Penn’s Kagan says. For example, last November Deere & Co. showed off its upcoming, fully autonomous 8R tractor. But as Wired magazine reported, Deere plans to collect data from the 8R just as it has for years with other AI-enhanced farm equipment.
The data help train new algorithms in order to develop future products and update equipment in the field—which Deere says benefits farmers. But some farming advocates say the data-harvesting could make Deere product owners overly reliant on the company. In a November University of Florida blog post, Ziwen Yu, assistant professor of agricultural and biological engineering, urges farmers to aggressively assert their rights to the data. It could turn out to be as or even more valuable than the crops, he emphasizes, because ag tech companies use it to create other high-value products. “The bottom line for this evolution,” he says, “is whoever owns the data … can claim the exclusive right to license [their] access and use.”
Those questions are still to be resolved, as are other reasons behind farmers’ reticence. Since many growers simply want to see the technologies work as promised, AgAID and AIFARMS are planning demonstration farms to showcase innovations. Investigators also are listening to farmers in order to deliver an invention that will get used. “We are letting them guide us, so we address what’s important to them,” says Kagan. AIFARMS and AgAID researchers, for instance, regularly meet with working groups of farmers interested in new technologies. “There are enough early adopters to help us get going,” insists Kalyanaraman. That should ensure that the seeds of a precision-agriculture revolution are sown in fertile soil.
Thomas K. Grose, Prism’s chief correspondent, is based in the UK.
Design by Francis Igot