The fruit you cannot see

Machine-vision yield estimation cannot count what the canopy hides, so it counts what it can see and multiplies by a correction factor calibrated on sample trees. This asks what that factor actually is, whether it is one number, and what it costs when it is not.

The canopy

A fruit is visible when the sight path to it is not blocked. Foliage extinguishes that path the way it extinguishes light, so the visible fraction follows Beer–Lambert in the leaf area between camera and fruit, integrated over where fruit actually sit.

Leaf area index 2.0

Extinction coefficient 0.55

Canopy variation across the block ±30%

Cameras

Extinction coefficients run about 0.3 for erect foliage to 1.0 for horizontally held leaves. Orchards are discontinuous canopies, where interception reaches an asymptote well short of the value a continuous Beer–Lambert canopy would give, so treat the numbers as the shape of the problem rather than a calibrated orchard model.

counted hidden from the camera

0% visibleblue band = measured across real orchards100%

63%of fruit visible to the camera

1.58×correction factor this canopy implies

±24%spread in that factor across the block, from canopy variation alone

One factor, two questions

The correction factor is calibrated by hand-counting representative trees. The published objection to that is not that it fails, but that choosing a representative tree is problematic given how much canopy density varies between trees. Here is what that variation does, depending on what you are trying to learn.

Block total ±1.2%

Any individual tree ±24% Error on a single tree's count, because that tree's canopy is not the average canopy. Thinning, irrigation and harvest scheduling are per-tree and per-zone decisions. They are made on the column that does not average out.

What this is. A first-order visibility model, not an orchard simulator. Visible fraction is the Beer–Lambert transmittance e^−k·LAI·d integrated over fruit depth d taken uniform through the canopy; two-camera visibility treats a fruit at depth d from one side as at 1−d from the other. Block error falls as 1/√N over the trees in the block. What anchors it. Single-side visibility measured across real apple orchards runs 40.85% to 79.83%, a range this model spans at k·LAI between about 0.5 and 2.2, which is the blue band above. Reported correction factors near 1.15 sit at the transparent end of that range. Companion to the grain bin and the soil core, which ask the same question where the barrier is heat and where it is statistics. Part of the Glass Lab × Charlot Lab track on Physical AI and integrated sensors for agriculture surveillance.