A plant-disease model that scored 99.35% on held-out lab images scored 31.4% on a separately collected set of online leaf images. The gap is not a bug: it is dataset shift. Lab datasets such as PlantVillage contain 54,306 single leaves photographed in controlled conditions against uniform backgrounds; field photos contain shadows, soil, clutter and multiple leaves. Below is what changes between the two worlds and the photo checklist, which has not been benchmarked here.
| Condition | Accuracy | What the images look like |
|---|---|---|
| Held-out lab images | 99.35% | One leaf, centred, plain background, even light |
| Real field photographs | 31.4% | Whole plant, soil and clutter, variable light, camera shake |
Sources: The Leaf and the Machine and the datasets it documents — PlantVillage (54,306 images, 14 crop species, 38 classes), iNaturalist's 2018 set (859,000 images, 5,000+ species), Leafsnap and PlantDoc.
Why does plant disease detection accuracy drop in the field?
Because accuracy is a property of the image distribution, not of the model alone. Four things change between a lab capture and a garden photo:
- Background. A model can learn "plain grey background = healthy" without learning the disease. Change the background and the signal disappears.
- Framing. Lab sets are cropped single leaves; field photos are whole plants. Lesion area becomes a few percent of pixels instead of most of them.
- Light. Colour is the primary disease signal, and colour shifts with white balance, shade and time of day.
- Class imbalance in the wild. Real gardens contain many diseases the training set never saw, so the model must answer "unknown" far more often than any fixed class list allows.
This is why a headline accuracy figure tells you almost nothing unless it states the world it was measured in — the closing argument of the film, and the design constraint behind our own on-device identifier.
What photo should I actually take for a plant diagnosis?
A repeatable protocol gets most of the field gap back. Shoot in this order:
- One leaf, filling the frame. Get within 20-30 cm; avoid the whole-plant shot until you have a close-up.
- Diffuse daylight, no flash. A north window or open shade. Direct sun clips highlights; flash flattens the colour the model reads.
- Plain background. Hold the leaf against a wall, a sheet of paper, or your palm rather than the soil surface.
- Photograph the symptom AND a healthy leaf. Comparison lets you and the tool see the difference, not just the lesion.
- Include the underside for suspected pests or mildew.
Then treat the result as a ranked hypothesis, not a verdict — confirm against the plant's own symptoms and care guide. If the leaves are already yellow, start with yellow leaf troubleshooting; for spots and rot, the encyclopedia's "Common problems" section lists documented causes per species.
How does the AI Planter identifier handle this?
It runs entirely on your device in WebAssembly — photos never leave the browser — and offers the following checks. These design choices do not establish its real-world accuracy:
- It matches against a visual-profile + embedding model of 282 species rather than a fixed disease-class list, but a ranked list can still be wrong, including when a plant is outside that reference set.
- It flags low confidence and non-plant images ("no green coverage") but those checks can miss unfamiliar cases.
- Every result links to that species' care guide, so the output is checkable.
Test it with your own field photos: identify a plant from a photo →.
Method and limitations
- The 99.35% / 31.4% figures come from Mohanty, Hughes and Salathé (2016): held-out PlantVillage images versus a separate collection from online sources. They are not measurements of AI Planter or of this page’s photo checklist.
- Dataset sizes (54,306 PlantVillage images; 859,000 iNaturalist images) are the published counts listed in the film's source list.
- What this page adds is the workflow: a five-step capture protocol and an explanation of which distribution shift each step addresses.
- Not a benchmark. We have not field-tested third-party apps; the accuracy figures describe the published lab-versus-field comparison, not a ranking of products.