Photographed in the wild, not on a lab bench — and held out from every split the models were fitted on.
It called the crop grape at 70%, and the grape disease model then said black rot at 100% — which is what the leaf actually has.
grape70%
the plant really is grape
black rot100%
clears its floor, so it is shown
black rot
what the people who built the corpus wrote down
How the app decides, in three steps
A crop classifier decides what plant it is looking at, then that crop’s disease model runs — and only speaks if it clears its own confidence floor.
The floor is the part that matters. A model that always answers is confidently wrong on everything outside its training; one that must clear a bar can decline instead.
More crops, same chain
Held-out field photographs, with the answer the corpus recorded. The chain named the crop, named the condition, and was confident enough to show it.
strawberry99%
the plant really is strawberry
healthy100%
clears its floor, so it is shown
healthy
what the people who built the corpus wrote down
potato91%
the plant really is potato
late blight100%
clears its floor, so it is shown
late blight
what the people who built the corpus wrote down
apple94%
the plant really is apple
rust99%
clears its floor, so it is shown
rust
what the people who built the corpus wrote down
How often it is right when it speaks
What matters is not how often a model is right, but how often the answers it puts on the screen are. Of 361 held-out field photographs, 131 were held back by a confidence floor and shown nothing.
| crop | test images | answered on screen | of those, right | precision | 95% CI |
|---|---|---|---|---|---|
| grape | 24 | 16 | 14 | 88% | 64%–97% |
| strawberry | 11 | 7 | 6 | 86% | 49%–97% |
| potato | 25 | 15 | 11 | 73% | 48%–89% |
| pepper | 24 | 11 | 8 | 73% | 43%–90% |
| apple | 59 | 38 | 25 | 66% | 50%–79% |
| tomato | 123 | 61 | 26 | 43% | 31%–55% |
| corn | 72 | 65 | 20 | 31% | 21%–43% |
| cherry | 10 | 6 | 1 | 17% | 3%–56% |
| peach | 13 | 11 | 0 | 0% | 0%–26% |
The faded rows are the crops this page does not showcase. They are here because leaving them out would make the four above look like the whole picture.
What this page does not claim
- Not a validated diagnosis. Nothing here is cleared for field use. These numbers motivate the gate; they are not a release.
- Small numbers where they are small. Two crops have fewer than ten shown answers, which is why every rate carries an interval.
- One corpus. A second corpus, season and set of phones would each move these numbers.
- Correct means matching a recorded label, not a pathologist re-examining the leaf. Where the label is wrong, this page inherits it.
Rebuilt with python scripts/render_crops_demo.py; the
measurement is scripts/measure_covered_crops.py.