Plant Disease Detection by AI: The Real Accuracy

Search for anything on AI-based plant disease detection and you reach striking numbers quickly: 99 percent accuracy, 98 percent, sometimes higher. Those figures are real and published in credible venues. But if you start a project in an actual orchard on the strength of them, the result you get will almost certainly be much lower.

The reason is neither a weakness in the technology nor an error in those papers. It is that the number was produced under conditions that differ from the conditions in your orchard. Understanding that difference is the most important thing to know before starting such a project — and it is the subject of this article.

Alborz province, with its orchards and greenhouses, and Karaj as a centre close to the surrounding farmland, is a suitable environment for this kind of work — because repeated on-site imaging, which we will see is the main condition for success, stays practical when the technical team is in the same province. We set out the wider context for these services in AI in Karaj. Which is precisely why starting from the wrong expectation is expensive.

Where the 99 percent comes from

Most plant disease models are trained and evaluated on standard datasets. The best known of them consists of individual leaves, detached, placed against a uniform background and photographed under controlled light.

Research using an EfficientNetB0 architecture on such data reported 99.51 percent accuracy. That number is correct — for those conditions.

Broader reviews of the same field carry a more uncomfortable finding: models reaching above 95 percent on these datasets degrade noticeably under field conditions. Three causes are cited: dataset bias, domain shift, and the complexity of several diseases appearing at once.

The point that deserves the most attention

There is a finding in those reviews more important than the accuracy drop itself: models trained on these datasets frequently rely on background features rather than the disease symptoms themselves.

The model has learned that diseased images in that collection have a slightly different background or lighting, and it classifies on that. It performs correctly on the test set too, because the test set comes from the same distribution. But in your orchard, where the background is soil and weeds and branch shadow, that cue no longer exists and the model has nothing left to lean on.

This is exactly why our computer vision page says a model must be trained on data from the same environment. A model that has never seen your light, your angles and your background does not work in practice — whatever number is printed in its paper.

What is a realistic figure?

When evaluation runs on data closer to real conditions, the numbers get lower but more dependable. Recent research with newer architectures reports 93.1 percent accuracy on training and 92.5 percent on test data. Transformer-based models have also shown better performance than conventional convolutional networks at detecting subtler disease patterns.

Ninety-two percent against ninety-nine may look like a decline, but in practice it is the opposite: a number produced under real conditions is something you can plan against. A number produced in the lab only shows that the method works.

Does that mean the project is not viable?

No. It means the project has to start in the right place.

Early image-based disease detection has clear practical value: inspecting every tree in an orchard by hand takes time, and symptoms are usually noticed only once the disease has spread. A system that surveys the whole orchard daily at ninety percent accuracy will, even with that ten percent error, flag a problem sooner than a weekly human inspection does.

The key is that the system must be designed as a screening tool, not as the final diagnostic authority. Its output is a list of points to inspect; the final judgement belongs to an agronomist. Designed this way, the model's ten percent error becomes a few extra inspections rather than a wrong decision.

The same logic — start narrow, measure honestly, keep a human in the loop — underpins the two other agricultural applications we have looked at separately: precision irrigation and water productivity and predictive greenhouse control.

A roadmap for a real project

Step one: pick one disease and one crop

A project that tries to detect every disease on every crop fails. The right start is one specific disease on one specific crop — ideally the one causing the most loss and showing clear visual symptoms.

Step two: collect images from the site itself

This is the most laborious and the most important part. Images must be taken in that orchard, with that camera, at different hours and in different weather. A few hundred real images from the site are worth more than tens of thousands of laboratory ones.

Step three: label with an expert

Images must be labelled by someone who knows the disease. Model accuracy never exceeds label accuracy, and this is the point at which projects most often cut corners.

Step four: evaluate on held-out data

The model must be evaluated on images it has never seen, ideally from a different season or a different plot. Evaluating on data that shares a root with the training set is exactly the mistake that produces the 99 percent figure.

Step five: build internal capability

Orchard conditions change from year to year and the model must be retrained on new data. If only the contractor can do that, the system becomes inaccurate and is abandoned within two seasons. An AI training course matched to the team, alongside an AI consultant for education, makes that independence possible. We set out the full approach in enterprise AI training.

Step six: expand gradually

Scope widens only after the result is proven on one disease and one crop. Every addition needs its own evaluation.

Frequently asked questions

How accurate is AI plant disease detection in reality?

Under laboratory conditions on standard datasets, accuracies above 99 percent have been reported. But scientific reviews show this degrades noticeably in the field; figures closer to 92–93 percent appear in more realistic evaluations.

Why does the model perform worse in the field?

Three main reasons: dataset bias, differing imaging conditions, and several diseases present at once. More importantly, research has shown some models rely on background features rather than disease symptoms — features that do not exist in the field.

How many images are needed to start?

There is no fixed number; it depends on how variable the conditions are. A few hundred real images from the site itself is usually an acceptable starting point. Quality and variety matter more than volume.

Is a special camera or drone required?

Usually not, to begin with. Many projects start with a phone camera or a simple fixed camera. A drone becomes worthwhile when the area is large enough that walking it for photography is impractical.

What is the final output?

A list of suspect points to inspect, not a definitive diagnosis. The system is designed as a screening tool and the final judgement stays with the expert; that way the model's error becomes an extra inspection rather than a wrong decision.

Computer vision for image-based screening, trained on data from your own site
Data science for analysing spread patterns and prioritising inspection
AI consulting for business to assess feasibility before any spending
AI automation for running the monitoring cycle and producing reports
Robotics and automated field imaging systems
Chatbot services for relaying notifications and guiding operators
AI training course for the technical teams of agricultural operations
AI consultant for education to design the internal team's learning path

Qatreh is based at Alborz Science and Technology Park and offers these services as a complement to organisations' existing programmes.

Sources

Crop disease detection using EfficientNetB0 deep learning approach for precision agriculture — Discover Computing, Springer
A comprehensive review on AI-based crop disease detection using leaf image classification and explainable AI — Discover Applied Sciences, Springer
Empowering Agriculture from Pixels to Diagnosis: A Review of Computer Vision Techniques for Plant Disease Detection — Archives of Computational Methods in Engineering