Read a leaf.
Photograph a single leaf. Apollo returns the crop, the condition, and how far it trusts its own answer. When the image is outside what it can read, it says so rather than guessing.
Or try one
The reading appears here.
What makes a readable photo
Accuracy in the field is mostly a photography problem. Four habits carry most of it.
01
One leaf, not a field
The model reads a single leaf. An aerial shot of rows gives it nothing to focus on.
02
Fill the frame
Hold the leaf close enough that it covers most of the picture. Detail is what carries the diagnosis.
03
Plain background
A hand, a sheet of paper, bare soil. Clutter behind the leaf pulls the prediction around.
04
Even daylight
Open shade beats direct sun. Hard shadows read as lesions that are not there.
How it's built
A ResNet-18 classifier trained on PlantVillage, 70k+ images across 38 leaf conditions, exported to ONNX.
That corpus is lab photography, so the network has never seen a tractor or a sky. Rather than let it force those into one of 38 diseases, Apollo scores the strength of the evidence behind each prediction and declines the ones it cannot support.