Identifying a sick plant with your phone
A model that was 99 % right on its own images dropped to 31 % on photos taken elsewhere. With the right photos, an app answers better.
A photo of a spotted leaf is rarely enough for a diagnosis. In a trial published in 2016, a model that was 99.35 % right on its own images dropped to 31.4 % on photos taken in other conditions. The lesson still holds. The app sees the spot, not the rest of the garden. A few well-taken photos help it answer better.
Photo diagnosis, according to the studies
Several apps offer a diagnosis from a photo. A 2022 study searched for them in three app stores. Out of 606 results, 17 met its criteria. Most had a decent score for ease of use, but a low one for analysis by artificial intelligence. The authors did not measure how accurate the diagnoses were. They say so themselves.
We found no independent study giving the rate of correct answers of these apps on vegetable garden photos. The most solid figure is still the 2016 one.
An app recognises photos close to its training images well. Your leaf, in your light and against your garden background, does not always look like them.
Four clues missing from the photo
A disease is recognised as much by its context as by its spot. Yet a photo isolates one leaf. It says nothing about these four clues.
- Where the spot appeared first. Early blight and septoria leaf spot of tomato start on the lower leaves, according to the University of Wisconsin.
- How fast it spreads. In warm, humid weather, tomato and potato blight spreads fast, according to the RHS, the British horticultural society.
- The underside of the leaf. In humid weather, blight forms a fine white fuzz there. Lettuce downy mildew does the same, under discoloured patches.
- The weather of the past days. Blight needs several hours of wet foliage to infect the plant.
You know these clues or can check them in two minutes. The app cannot.
Five common problems to check in the garden
| Problem | On the photo | Also check |
|---|---|---|
| Blight (tomato, potato) | A watery rot that spreads, leaves that turn brown and collapse | The underside of the leaf on a humid morning, with a fine white fuzz. And the rain of the last few days |
| Early blight (tomato) | Round brown spots, with rings like a target and often a yellow halo | It starts on the lower leaves |
| Septoria leaf spot (tomato) | Small spots of about 6 mm, with a whitish centre and a dark edge | It also starts at the bottom. The spots end up merging |
| Powdery mildew (courgette) | A white powdery felt on the leaves | Irregular watering, a plant against a wall where the air does not move |
| Lettuce downy mildew | Discoloured patches on the upper side, often bounded by the veins | A white or grey mould on the underside, after a long spell of humid weather |
| The move | Before taking the photo | Turn a leaf over, look at the base of the plant, think back to the week's weather |
Sources: RHS for blight, powdery mildew and lettuce downy mildew. University of Wisconsin for early blight and septoria leaf spot.
On a blurry photo, early blight and septoria leaf spot are easy to confuse. Both make brown spots that start at the bottom. On a sharp photo, the size of the spots and the rings help you decide. The guide Tomato blight explains how to tell blight from early blight. It also shows how to read the weather to stay ahead of it.
Photographing a diseased leaf well
A 2023 study on plant identification looked at why apps succeed or fail. The sharpness and exposure of the photo did not change the result significantly. The framing did. For some apps, a plant well isolated in the image gave better answers. The study covered wild species, not diseases. In the vegetable garden, the principle holds. Show the diseased part clearly.
Then read the next suggestions, not only the first. In the 2023 study, all apps combined, the first suggestion was right for 69 % of the photos. The right answer was among the first five for 85 % of them.
If the clues contradict each other, ask for advice. A garden centre or an experienced gardener can often recognise the disease by eye. Bring an affected leaf in a closed bag, so you do not spread spores on the way.
The test we have not done yet
We wanted to submit ten photos of known problems to several apps and publish the rate of correct answers. This test is not done. Until it is, we give no figures of our own.
You can do it at home in ten minutes. Take three photos of a problem you already know, for example powdery mildew on a courgette at the end of summer. Submit them to two apps. Note whether the right answer comes first, in the first five or not at all. You will soon see whether an app copes with your own photos.
Willy, in the Vegiwise app
Willy, our assistant in the Vegiwise app, accepts up to three photos per message. He reads them together with the plants in your garden, the weather at your address and the current alerts. He is an artificial intelligence. The app says so itself, he can make mistakes. Show him the underside of the leaf. Tell him when the spot appeared. His answer will be better. To see him at work, go here.
To choose a vegetable garden app more broadly, read Choosing a vegetable garden app that knows your garden.
FAQ
Does an identification app also recognise diseases?
Some do. Their strong point is still the species. In the 2023 study, the right species was among the first five suggestions for 85 % of the photos. No comparable study exists yet for vegetable garden diseases.
Why do two apps give two different answers?
They did not learn from the same photos. According to the 2016 study, a model loses a lot of accuracy when the photo does not look like its training photos.
Should you treat as soon as an app sees a disease?
No. First confirm with the underside of the leaf and its position on the plant. Several steps come before any product. Remove the affected leaves, let the air through, water at the base. Before using any product, check that it is approved for amateur gardeners in your country.
Can an app get the plant wrong?
Yes. In the 2023 study, none of the five apps always found the right species in its first answer. The average was 69 %. For a plant you sowed yourself, you know the answer. Tell the app if it lets you.
Photo: Wolan268, Wikimedia Commons, CC0
Sources
- Mohanty S. P., Hughes D. P. et Salathé M., Using Deep Learning for Image-Based Plant Disease Detection, Frontiers in Plant Science 7, 1419, 2016 (54 306 images, 14 cultures, 26 maladies)
- Siddiqua A. et al., Evaluating Plant Disease Detection Mobile Applications: Quality and Limitations, Agronomy 12 (8), 1869, 2022 (606 applications trouvées, 17 évaluées)
- Hart A. G. et al., Assessing the accuracy of free automated plant identification applications, People and Nature 5 (3), 2023 (857 photos, 5 applications)
- RHS, Tomato blight (pourriture aqueuse des feuilles, feutrage blanc au revers, plusieurs heures de feuillage mouillé)
- RHS, Potato blight (Phytophthora infestans, le même organisme que sur la tomate)
- University of Wisconsin Extension, Early blight (taches en cible, halo jaune, feuilles du bas d'abord)
- University of Wisconsin Extension, Septoria leaf spot (taches d'environ 6 mm, centre blanchâtre, bord foncé)
- RHS, Powdery mildews (feutre blanc poudreux, courgettes, stress hydrique irrégulier, air immobile)
- RHS, Downy mildews (taches sur le dessus, moisissure au revers, laitue, feuillage mouillé longtemps)
He gardens in Belgium and tries out in his own vegetable patch what the app recommends. His articles always start from a source you can check and a step you can take this weekend.
