The Decision Satellites Can't See
Kristin Scholten is a farmer turned data scientist in the advanced analytics department at Nationwide, and she is building OpenFields, an AI prediction engine for agriculture. At the H3 Conference 2026 in New York City, she argued that crop acreage is decided months before any satellite can see a plant, and showed how she used H3 to put a crop probability on every field in North Dakota.
Acreage moves the market
A couple of years ago, Kristin gave a talk on estimating crop yield from satellites using solar-induced fluorescence, which she calls a heart rate monitor for plants. Afterward, an agricultural commodity trader asked: can you do this for acreage? Acreage causes bigger swings in the market, and it is better to know it early.
According to Kristin, the USDA misses its acreage number by about 5 million acres every year. That is 582 cells at H3 resolution 6, roughly the size of New Jersey, and about $3 to $4 billion in the market.
Satellites see the crop too late to help. The USDA publishes projections in March, farmers plant in April and May, projections are updated in June, harvest runs through October and November, and the final numbers come out in January. The planting decision itself is made the previous fall, before there is anything to photograph.
A farmer's choice comes down to local prices
To make the point, Kristin invited an audience member to run a field in North Dakota and pick a crop for 2027. The field's history since 2018 followed a corn and soybean rotation that pointed to corn. Then she added information a farmer actually weighs:
- Input costs. Corn needs a lot of fertilizer, soybeans need little, and spring wheat sits in between.
- National prices. December 2027 futures were $5.26 per bushel for corn, $12.87 for soybeans and $7.45 for spring wheat.
- Local basis. The grain elevator, where farmers sell their harvest, adjusts prices based on its own supply. After the discount, the elevator paid $4.31 for corn, $12 for soybeans and $7 for spring wheat.
With national prices and the rotation, corn looked like the answer. With the local elevator price, soybeans came out ahead, and the volunteer switched. That flip is what Kristin wants to predict.
From hexagons back to fields
H3 is how she gets a local price for every field. Using elevator locations, and the cost of hauling grain to them (some farms can choose between two elevators), she calculates a price for each hexagon. That value is then assigned to the fields inside each hex, which lets her estimate, field by field, the probability that a farmer flips their rotation. A single field might come out as a 52% chance of corn and a 41% chance of soybeans.
Asked why not compute distances point to point, Kristin said that would be the most granular approach but too compute heavy. Each hex holds hundreds of fields, so aggregating at the hex level and assigning values back to fields keeps the analysis manageable. The usual alternative, county-level aggregation, is not granular enough.
900,000 fields and counting
Kristin ran the model for all of North Dakota: more than 900,000 fields, predicting the 2026 growing season that was weeks from harvest at the time of the talk. The field boundaries are open data, much of it from the USDA, in line with the OpenFields mission of sharing data openly.
The fall model updates about once a month. In the spring it updates weekly, using factors such as frost-free days: North Dakota is getting more of them, so more farmers can grow corn, but a wet spring pushes acreage toward soybeans, which have a shorter growing season.
Her next step is measuring accuracy, including checking this season's predictions against satellite observations, and she invited collaborators to help.