Artificial Intelligence for Agriculture Needs Readable Data

USDA Asked for AI That Can Read the Plant Data It Already Collects

On July 22, USDA called on universities and other partners to build AI tools that can translate its germplasm data: the images, field records, and lab results the agency has accumulated from the seeds and plant material it keeps. The tools would pull those three kinds of records together so scientists can identify plant and seed traits faster and breed more resilient crops.

Notice what the request concedes. The constraint USDA describes sits upstream of model quality, sensor coverage, and compute: the archive is bigger than the agency's ability to mine it.

Two federal actions on AI and agriculture landed inside the same seven days, and a cluster of research papers landed next to them. All of it points at the same bottleneck: whether the records an organization already owns can be read by a machine.

The request, and what it quietly admits

USDA's ask runs through the Genesis Mission, the White House Office of Science and Technology Policy effort launched in November 2025 to aim AI at scientific discovery, and through the Agriculture Advanced Research and Development Authority, or AgARDA. An Agricultural National Science & Technology Challenge is expected to launch later in 2026. Teams that take part would get access to the American Science and Security Platform, shared infrastructure built by the Department of Energy that connects datasets, instruments, and AI tools in one place.

"USDA is taking concrete steps to give scientists the modern tools they need to innovate agricultural solutions from the vast plant data that they collect," said Dr. Scott Hutchins, REE Under Secretary and USDA Chief Scientist.

A germplasm collection is a hard target for AI for reasons that have very little to do with the models. It is not one dataset. It is decades of photographs, field observations, and lab assays captured under different protocols, by different people, for different projects, with much of the context living in a lab notebook or in the memory of whoever ran the trial. Some of it is digital and some of it is not. The formats change, the naming changes, the units change.

None of that means USDA scientists cannot use the collection. They use it constantly. It means the pace of that work is set by how readable the records are, and the agency has now said so in public, in a request for outside help.

That composition should sound familiar, because it is roughly the composition of every operational archive in agriculture. Neither report on the announcement discloses how large the collection is or how many records sit inside it, which is a small piece of evidence in its own right. For anyone weighing whether the next AI investment should be a better model or a data cleanup, the agency that runs the country's plant genetic resources has published its answer.

Congress wants the same bet written into statute

In the same stretch of days, the legislative version of the same idea picked up a second chamber. Rep. Don Davis introduced the House companion to the FARM AI Act, formally the Fostering Agricultural Research and Modernization through Artificial Intelligence Act, cosponsored by Rep. Zach Nunn. Sen. Ted Budd introduced the Senate version back on March 21.

The bill would make artificial intelligence for agriculture a priority research area funded through the Agriculture and Food Research Initiative and AgARDA, give USDA Extension the resources to help farmers adopt AI and precision agriculture, expand grants and fellowships aimed at the rural workforce, and require USDA to designate a senior official as its AI in agriculture advisor to align federal programs.

It would also direct USDA to work with the National Institute of Standards and Technology on national standards for agricultural AI. That provision has the longest half-life of anything in the bill. Standards work is how "we used AI" turns into a specific claim: this system, on these inputs, evaluated this way. No such rules are drafted, and the bill is introduced rather than passed, so anyone predicting their content is guessing.

Extension is the piece most growers would actually touch. If the adoption money lands there instead of in grants to vendors and universities, the practical effect on a mid-size operation looks very different, and which one the final text funds is the detail to follow.

Anyone who has watched federal AI standards work in other sectors knows the shape of the eventual questions, and they are narrow ones: what the system did, on which inputs, and who approved the output. Those are record-keeping questions before they are modeling questions, and a record is easier to keep as you go than to reconstruct two seasons later. On AgentPMT, an agent's every tool call is written to an audit log down to the request and the response, so the history of what ran on which data exists before anyone asks to see it.

Budd made his own case for the bill in a Washington Times op-ed on July 26, arguing for folding its provisions into the current Farm Bill and framing the technology as something that extends the capacity of farmers and the rural workforce rather than replacing either. That is the sponsor's advocacy for his own legislation, not reporting, and it reads accordingly.

Four preprints, one week, the same premise

While Washington was describing the problem, four preprints posted between July 20 and July 24 quietly demonstrated it. Every one of them pulled new predictions out of data agriculture had already collected. All four are preprints and have not been peer reviewed.

The sharpest of them forecasts sweet pepper harvests. A team led by Enrico Pallotta built a multimodal model over an image time series and per-plant fruit counts from two greenhouse growing seasons, then measured it against a persistence baseline, which is the working assumption that next week's harvest resembles this week's. The model cut forecast error by 33% for the 2022 season and 38% for 2023. For a grower, that is the difference between staffing and selling against a guess and staffing against a number, produced from photographs and counts the greenhouse was already generating. The model does no automated harvesting and drives no machinery. It tells you how many fruits will be ready to pick.

The second traces field boundaries. Researchers ran a residual U-Net over 1-meter imagery from the National Agriculture Imagery Program, refined the difficult cases with a text-prompted segmentation model, and produced a farmland extent layer accurate enough to be usable where field boundary data simply does not exist. NAIP has been flying that imagery for years.

The third listens to beehives. Recurrent networks trained on the public UrBAN dataset, more than 3,000 hours of hive audio, estimate colony strength by preserving the time dimension of the sound that earlier feature-engineering approaches threw away. The recordings were already public. What changed was the representation.

The fourth predicts sugar beet yields from Sentinel-2 satellite bands, flagging a large share of low-yield fields early in the growth cycle. Its most useful finding is a design one: very small vision transformer patch sizes and the full set of spectral bands both improved results, and both are uncommon choices in remote sensing practice.

Separately, a Molecular Plant paper published July 22 by ICRISAT with IPK Gatersleben and the University of Queensland argues for pairing genebank diversity with AI prediction models and functional genomics. It is a strategy argument rather than an experiment, so no measured outcome belongs to it, but it stakes out USDA's premise from the research side: the underused input is the material already conserved and characterized.

Reading these four together as one week's verdict is our synthesis, not a claim any of the authors made. They are separate groups working separate crops on separate continents. What links them is what none of them needed. No new satellites. No new sensor packages. No new instrumentation of any kind. Every project supplied the labeling, formatting, and modeling that made existing records tractable, and the predictions fell out of data that had been sitting untouched for years.

The agriculture automation story usually opens with equipment: a new implement, a new sensor rig, automated farming equipment priced like a second mortgage. This week's results came from a much cheaper place.

The same gap sits in your own records

USDA's three-part phrasing, images and field data and lab results, describes a farm, a co-op, or a processing plant about as well as it describes a federal genebank. Photos on somebody's phone. Scouting notes dictated into a recorder or written on the back of a work order. Equipment and irrigation logs in whatever the manufacturer's portal exports. Lab and QA results as PDFs, or on paper in a filing cabinet nobody has opened since the last audit.

The Challenge, whenever it opens, will fund research institutions. It is not coming to your operation, and neither is the FARM AI Act's Extension money for a while yet. So the useful move is the one that runs on your own schedule: inventory what records exist, mark which are machine-readable today, and identify which are trapped in formats no model can touch. That last category is usually the biggest and always the least glamorous, and it is entirely under your control.

Each of USDA's three categories has an ordinary answer. Spoken field observations, the agronomist-with-a-recorder version of "field data," can be landed as structured rows instead of audio nobody replays; AgentPMT ships exactly that as a workflow, Plaud Spoken Field Notes to a Structured Sheet. Lab and QA reports stuck in PDFs and paper are what a document OCR agent exists for, turning scans into records a query can reach. And the public reference series that field-level questions always seem to need are available to an agent as callable tools, so it can join an operation's own numbers against agricultural and food security data or climate, environment, and land data without a custom integration for each source. That is the shape of AgentPMT's agriculture and food production work: the unremarkable translation jobs that sit between a shoebox of records and a model that can use them.

None of this substitutes for the federal effort. It is the same translation work USDA is asking universities to take on, at the scale of one operation, and it has the advantage of being finishable this quarter. The capability is not the scarce part, which is a pattern that shows up well outside agriculture: the intelligence keeps arriving faster than the systems that feed it.

The Challenge has not launched. The FARM AI Act is a bill with a companion, not a law, and the standards it would produce do not exist on paper yet. Both could stall. What the week's papers showed does not depend on either: agricultural records that have been sitting unread still hold predictions worth having, and getting at them requires no appropriation, no rulemaking, and no waiting. The archive is already yours. Somebody has to make it readable.

---

Sources

• USDA Asks Partners to Develop AI Solutions to Accelerate Crop Innovation, RFD-TV
• USDA seeks AI partners to accelerate crop innovation, Blue Book Services
• New FARM AI Act to expand AI research, training at USDA, Carolina Journal
• Put AI to work for American farmers, The Washington Times
• Forecasting the Number of Harvest-ready Fruits of Sweet Peppers Using Multimodal Time-Series Data, arXiv
• Farmland Extent and Visible Boundary Mapping from 1 m NAIP Imagery Using Residual U-Net and Text-Prompted SAM 3 Refinement, arXiv
• Improved Monitoring of Honey bee Colony Strength via Audio IoT Sensors, Modulation Tensorgrams and Recurrent Neural Networks, arXiv
• Early Yield Prediction for Sugar Beet Fields using Satellite Data, arXiv
• What happens when Pangenetics meets AI-powered Prediction? Next Frontier of AI-driven Crop Design, Global Agriculture