Why we built this
In the summer of 2021, Daniel Ferreira was consulting on a 2,300-acre corn operation in central Iowa during one of the driest Julys on record. The grower had a soil moisture sensor network - six nodes across the whole farm - and USDA county-level yield forecasts that were three weeks stale by the time they arrived. Nobody could tell him whether the southwest quarter was going to make 180 bushels or 140.
Daniel had been working with satellite imagery in remote sensing research before moving into agronomy consulting. He knew Sentinel-2 had the temporal resolution and the red-edge band to detect canopy stress weeks before it showed up in the crop. He built a rough pipeline over that winter, ran it on archived data against the 2021 weigh tickets, and found the mean absolute error was 4.3 bu/ac at the parcel level - better than anything available at a field scale.
The problem was that pipeline was a collection of Python scripts, a cron job, and a Google Sheet. It worked for one consultant who understood what he had built. It was useless for anyone else.
CropKern is what happens when you take that pipeline seriously. We spent 18 months making it reproducible, documented, and operable by agronomists who do not want to run Python scripts. The satellite and sensor fusion is the same. The forecast methodology is the same. The interface is not.
