Orchard yield estimation: age, density, canopy and past harvests

Counting trees is not estimating yield

The inventory tells you how many living trees there are and how big they are. Kilos are a different matter: they depend on the age of the planting, on management, on that year's water and on the crop's own alternate bearing. Between the count and the harvest there is a model, and every model has assumptions.

Our position is that those assumptions get published. A yield number that does not come with the formula, the source of each input and an error margin is not an estimate: it is an opinion with decimals. What follows is the whole model, exactly as it runs.

The model, in one line

Kilos = trees × kilos per tree at peak × age ramp × calibration factor. Four pieces, none of them hidden.

Kilos per tree at peak come from spreading a per-hectare yield over the typical density of the spacing: 22 tonnes per hectare over 333 trees is 66 kg per tree in citrus; 12 tonnes over 204 trees is 59 kg in avocado. Those are the crop profile defaults, the starting point when the plot has no history yet.

The ramp acknowledges that a young tree does not produce like a mature one: zero before the first productive year (the third, in citrus and avocado), and from there a straight line up to peak over five years in citrus and six in avocado. A five-year-old block sits at 60 % of its potential; a seven-year-old one is already at 100 %.

Calibration is what turns this into your orchard rather than a regional average, and it deserves its own section.

Calibration: your harvests outrank the average

As soon as you register a real harvest for a block, the model computes what it would have predicted for that year and takes the ratio: if you harvested 1.2 times the prediction, that block is corrected by 1.2 going forward. It is the most valuable part of the system and the cheapest to feed: one number per block per season.

The factor is bounded between 0.3 and 3.0, and that bound is deliberate. A ratio of 0.1 or of 8 almost never means the model is wrong: it means the observed figure is a partial harvest, two blocks recorded as one, or kilos captured in a different unit. Outside that range the figure is ignored, a warning is raised, and the estimate falls back to the model. A system that lets any number calibrate it learns garbage.

The effect on the margin is direct: with no harvests on record the estimate ships at ±30 %; with the block calibrated, at ±15 %. That margin travels with the figure everywhere, and it is not decorative: it is the difference between planning with a range and believing a decimal.

The area mistake: how a plot ends up yielding 700 tonnes per hectare

This deserves its own section because we see it at every new client's kickoff. Yield per hectare is a division, and the problem is almost never in the numerator: it is in which area you divide by.

There are at least three areas for the same plot, and they do not agree: the deed, the grower's registry and what the drone actually flew. In a real agave case, a plot the registry declared at 0.65 hectares measured 6.18 when flown: nearly ten times. Yield per hectare computed with the paper area came out in the hundreds of tonnes, an impossible figure that nevertheless propagated into reports until somebody looked at it.

The rule we apply, with no exception: the area you divide by is the flown one, the sum of the polygons actually covered. The registry area is shown beside it as a reference, and when the two differ by more than 50 % the figure ships flagged with an area alert instead of being printed clean. Try it in the calculator: enter a different paper area and watch what happens to the number you were about to present in a meeting.

What this model does NOT do yet

The tree version is version 1, and saying so is part of the delivery. Three things it does not do today, and which are the natural improvements once there are real harvests from several clients:

• It does not use canopy diameter. We measure it in the inventory, but for trees it does not enter the calculation yet; between two blocks of the same age, one with large canopies and one stunted, version 1 assigns the same potential.
• It does not use vigor. Per-plant NDRE exists when there was a multispectral flight, but it does not correct the estimate; today it shows where the problem is, it does not lower that block's kilos.
• It does not separate seasons within a year nor propagate one block's calibration to its neighbors, even when they share management and soil.

How the margin closes (the part that is up to you)

Almost all the remaining error is not the algorithm's: it is in the inputs. In order of how much they move the needle:

• Planting date per block. The heaviest variable in a young orchard. Without it the model assumes a mature block and says so, which is honest but also expensive: it can overestimate by a factor of two.
• Harvest recorded per block, not per plot. A plot total calibrates the average and hides exactly what matters, which is which block is underperforming.
• The same unit, always. Boxes, sacks and tonnes coexisting in one table is the number-one cause of out-of-range factors.
• A real count instead of assumed density. With no flight, the model multiplies area by the theoretical spacing; that ignores gaps, which run 3 to 10 % of the positions.
• Two seasons. With one harvest the model adjusts; with two it starts telling a block's good year from its bad one. That is how the yield estimation we deliver works.