A wrong call on irrigation, fertilizer, or pest control can cost farmers time, money, and entire harvests. But a new study suggests those decisions might soon be tested before they ever reach the field.
Researchers have developed a digital replica of a farm that can simulate real conditions and guide crop decisions with 87.5 percent accuracy in early trials.
The goal is simple: take some of the risk out of farming by turning uncertainty into something farmers can see and act on ahead of time.
Inside a simulated farm environment, streams from drones, ground sensors, and satellites merge into a single system that mirrors real field conditions.
By connecting those live signals to predicted outcomes, researchers at the University of Tennessee (UT) demonstrated that the system can match farmer requests to the correct actions while filtering out unsafe choices.
Performance held across varied test scenarios, where the system both applied appropriate rules and corrected violations before any decision reached the field.
Even so, those results define controlled conditions, leaving open how the same reliability will hold once real farms introduce unpredictable weather, pests, and soil variation.
Across corn fields, the fall armyworm, a fast-spreading crop-eating pest, has spread to more than 80 countries beyond its original range.
Meanwhile, agriculture uses about 70 percent of global freshwater withdrawals, making each irrigation mistake costly.
Excess fertilizer also escapes the soil because microbes transform spare nitrogen into gases or runoff that crops never use.
Early warnings matter more than broad averages when pests, drought, and nutrient losses can all change quickly
Above a field, drones collect hyperspectral imaging - pictures across many light bands - that can reveal crop stress before leaves visibly change.
When those scans flag trouble spots, people on the ground pull samples so quick tests can pinpoint the cause.
Results from air and soil then feed into the model, where current conditions are matched with likely next outcomes.
Farmers need that full chain, because a yellow patch could mean very different things - lack of water, missing nutrients, or the early stages of disease.
Inside the system's digital twin - a live virtual copy of the farm - software turns those signals into clear, usable rules. In 85 test cases, the framework matched plain-language requests to the right actions 87.5 percent of the time.
It also adds a layer of protection. Digital checks enforced safety limits in 88 percent of cases and corrected every detected violation before anything reached the field.
That filtering matters, because a bad recommendation could quietly break irrigation limits, spray schedules, or harvest rules before anyone notices.
Clear language matters because most growers need a decision, not a dashboard crowded with charts and color scales. Researchers use a large language model to translate sensor patterns into short recommendations.
Study lead author Charles Cao is an associate professor in UT's Department of Electrical Engineering and Computer Science.
"It's about putting practical tools in the hands of farmers around the world - tools that help them grow more food with fewer resources and less risk," said Cao.
Plain-language advice could widen access, especially for smaller farms that cannot hire technical specialists or data analysts.
Because the project spans four countries, the team built a federated learning, training shared models without moving raw farm data.
"Our approach creates an AI system that gets smarter from global experience without any single country or farm having to share its raw data," said Cao.
Local training matters because farm records can reveal land management, business strategy, and environmental patterns growers may not want exposed.
International teamwork also gets easier when privacy laws, weather patterns, and farming practices differ sharply.
Far from steady broadband, the group is building edge computing directly into drones and sensors, so data can be processed on the device itself instead of relying on remote servers.
That onboard processing allows aircraft to detect crop stress, pests, and soil conditions in real time.
Low-power radios then send small data packets across long distances, keeping systems updated even in remote areas. By reducing internet dependence, the approach could reach growers who are often left out of costly digital services.
The project is now moving beyond theory and into the field. The first sensor prototypes were set for deployment in spring 2026, marking a shift from lab testing to real-world use.
Within the broader effort, a Missouri-based engineering team is working alongside collaborators in Japan, India, and Australia. The National Science Foundation selected the project in February 2026 as one of six new initiatives.
The goals are ambitious: cut crop losses by 20 to 30 percent while also reducing water and fertilizer use by about 20 percent.
Even the strongest early numbers come from controlled tests, not from years of harvest results across commercial farms. Weather swings, local pests, crop varieties, and market pressure can still break neat digital plans once boots hit the soil.
Human judgment also stays central because a model can rank options but cannot own the risk of failure. Field trials now carry the burden of proving the promise beyond carefully designed test cases.
At the same time, Cao's project is designed to address many of those real-world gaps by combining earlier warning, safer automation, clearer advice, and protected data in one farm-scale system.
If field trials match the early results, farmers could practice decisions before making them and trim losses sooner.
-- -
Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates.
