Satellites Plus AI Could Soon Predict How Big Your Grocery Bill Will Be

Every year, the U.S. government publishes how much corn, soy, wheat, cotton, and sorghum each county grew. The trouble is that this number is one big average for a whole county. Modern AI wants to know what happened on each individual field, and that data basically doesn't exist. So a team of researchers built AgroBench, a free, public dataset that bridges that gap. They took official county harvest totals and matched them to satellite pictures of the specific farm plots inside those counties.

Think of it like a teacher who only knows the class average, not each student's score. AgroBench gives the AI thousands of individual "students" (crop fields) and a rough class average to learn from. The computer studies how the fields looked from space all season — color changes, plant growth, moisture — and learns to connect those patterns to the final harvest number. It uses European Space Agency satellite images, radar that sees through clouds, weather data, and maps of the land's shape and elevation.

The scale is big: over 13 million observations from 788,654 distinct crop plots, covering 5,107 county-year combinations from 2017 through 2024. Everything — the data, the recipe used to build it, and the testing rules — is being released publicly, so researchers and companies can compare their AI models fairly instead of each hiding behind private data.

Why should you care? Food prices swing wildly when harvests surprise us. Better forecasting means farmers can plan planting, fertilizer, and selling; insurers and grain buyers can price risk more accurately; and governments can spot trouble earlier. It's early-stage research, not a product, but it builds the shared foundation everyone else will stand on.