Measured against itself
The comparison is the whole design. Set a plot against its neighbours and you need a model to say what a comparable plot is, and every result then inherits the model’s assumptions. Set the same plot against its own 5 to 8 seasons and the comparison holds without either. It answers 2 questions cleanly: did this field do better or worse than it usually does, and did something change after a particular date.
The number of clear looks, printed every time
Cloud is the constraint nobody selling this product likes to mention. The long rains are the planting season and they are also the cloudiest part of the year, so a plot can go weeks without a usable satellite view. A figure built on 3 clear observations is not the same object as one built on 15. Pretending otherwise is how these studies go wrong.
So the count travels with the figure. Every greenness number we return is printed next to how many clear looks it rests on. A season too thin to carry a verdict is reported as thin rather than averaged into confidence.
Weather over that area, not over that plot
Rainfall and temperature come from public gridded products read at the plot’s own coordinates. Those grids are several kilometres across, which is why the wording matters: it is the weather over that area, and at a daily resolution it is better read as whether it rained than as how much. Monthly totals hold up well. A single day does not, and a claim that rests on one day is built on the weakest thing here.
You will not see “what the station recorded” anywhere in a report. Weather stations are a different product of ours, they sit on the farms that bought them, and a customer’s growers will almost never have one.
Yield, and the only defensible form of it
A per-farmer yield read from satellite is not a measurement at this field size, and we do not sell one. What does survive is comparative: a band saying how this plot did relative to its own record and relative to the rest of the cohort, plus an aggregate across a few hundred farmers where the individual noise cancels and the mean carries information. A programme evaluating an intervention across a district can use that. A lender pricing one farmer cannot, and should ask a different question.