How AI is helping greenhouse growers catch tomato diseases before they spread

It starts with a routine walkthrough of the tomato section. Everything looks fine from the aisle. Three days later, a grower spots the early signs of leaf mold spreading across a bench. By that point, the infection has already moved through a section of plants. Treatment begins, but by then it is damage control rather than prevention. The outbreak could have been detected earlier.

This scenario plays out in greenhouses more often than it should. Treatment begins, but by then it is damage control rather than prevention. The outbreak could have been detected earlier. Tomato crops are particularly vulnerable. Leaf mold, early blight, bacterial spot, and Septoria leaf spot can spread from a few infected leaves to a significant outbreak within 48 to 72 hours under high humidity.

The challenge is not that growers are not looking. It is that the human eye has limits, and disease does not wait for the next scouting round.

Why manual scouting has a built-in blind spot

Scouting happens on a schedule. Disease does not. A grower walking a greenhouse section every three to seven days will catch symptoms that are already visible.

The earliest indicators are harder to catch at walking pace: subtle discoloration, minor texture changes on the underside of a leaf, and faint spotting that precedes a full lesion. These early signs are also easy to misread, especially in large row systems where fatigue and time pressure affect the quality of observation.

Detection also varies between workers and shifts. One experienced scout may catch what a newer team member overlooks. In a high-volume greenhouse operation running multiple varieties across several structures, that inconsistency adds up. The result is that most disease responses begin after visible spread has already begun, making treatment reactive rather than targeted.

This is not a criticism of scouting. It is a structural limitation. Manual inspection was never designed to detect the earliest stages of tomato disease at the leaf level. That is a different kind of problem. It requires a different kind of tool.

How AI reads a leaf before you can

A computer vision model does not get tired or overlook a row, and it does not require symptoms to become obvious before flagging potential disease. It analyzes a leaf image for disease-specific patterns that are genuinely difficult to distinguish at early stages during a visual walkthrough.

Early blight presents as concentric ring-shaped lesions, Septoria leaf spot as smaller uniform spots with darker borders, and leaf mold as faint yellowing on the upper leaf surface before the characteristic gray-green growth beneath becomes visible. A trained model learns to distinguish these signatures from one another and from healthy tissue at a stage that most scouts moving at a normal pace will miss.

The Omdena team built and trained models specifically for tomato disease detection using Convolutional Neural Networks (CNNs), a class of machine learning models that are particularly effective at image recognition tasks. The models were trained on datasets of both healthy and diseased crop images, learning to distinguish between disease types based on visual patterns rather than hand-written rules.

Using transfer learning, the models were adapted from existing image-recognition systems and fine-tuned for tomato disease detection. CNN models trained on large plant disease datasets have reported accuracy above 90%, a benchmark the Omdena project sought to replicate under greenhouse conditions.

The practical output is a classification: which disease, and how confident the model is. That output serves as a prompt for targeted inspection, not a replacement for it. The AI flags the area. The grower makes the call.

What real deployments actually taught us

Building a model that works on a curated dataset is one thing. Building one that holds up in an actual greenhouse environment is another. The Omdena project surfaced several challenges that any grower considering these systems should understand before investing.

Lighting was the most consistent issue. Greenhouse lighting varies significantly across time of day, structure design, and season. A model trained primarily on outdoor or well-lit images can struggle with the shadows, reflections, and uneven illumination common in greenhouse settings.

Images taken in low light or with glare from overhead structures introduced errors that the model had not been trained to handle. In some cases, uneven lighting led to visually similar leaf conditions being misclassified, even when the same model performed reliably on images from the controlled dataset. This does not make the technology impractical, but it does mean image capture conditions are just as important as model performance.

Image quality and capture consistency were equally critical. Blurred images, extreme angles, and partial leaf coverage all reduced classification accuracy. The models performed reliably when images were captured close to the leaf, in reasonable light, with the affected area clearly in frame.

When those conditions were not met, performance dropped. Standardizing how images are taken, whether by staff protocol or fixed camera positioning, turned out to be as important as any technical refinement to the model.

The broader lesson is that the gap between a controlled training environment and a real greenhouse operation requires deliberate bridging. The technology works. How it is used determines whether it works consistently.

How to start without overcommitting

The most effective way to begin is not a full deployment. It is a focused pilot. Pick one greenhouse section growing a single tomato variety and one disease that has historically caused problems in that space. A limited pilot makes it easier to evaluate whether the system delivers meaningful value.

For image capture, a mobile device is sufficient to start. A grower or scout photographs leaves during regular walkthroughs, uploads the image to the system, and receives a disease classification alongside a confidence score.

A high-confidence result warrants immediate closer inspection of that section. A lower-confidence flag is still actionable, serving as a prompt to revisit that area within 24 hours rather than waiting for the next scheduled scouting round.

The workflow does not replace scouting. It changes where and when the scout pays attention. Fixed cameras can come later, once capture protocols are established and the team is comfortable with the outputs.

The most immediate operational benefit is localized treatment. When the disease is caught earlier and in a smaller area, the response can be targeted to that section rather than applied across a whole structure. That means lower chemical use, lower labor cost, and less stress on healthy plants from unnecessary treatment. Research in precision disease management suggests targeted early intervention can reduce fungicide applications by 20% to 40% compared to calendar-based spraying.

What this system cannot do (yet)

These systems are not plug-and-play. A model trained on one dataset will not automatically perform at the same level in a different greenhouse environment. Temperature, humidity, crop variety, and lighting conditions all influence how disease presents visually, and a model needs exposure to those conditions to classify reliably within them. Growers should expect a calibration period when starting, not immediate out-of-the-box accuracy.

Human validation remains essential. The AI does not make treatment decisions. It surfaces candidates for closer inspection. Growers who treat every flag as a confirmed diagnosis or dismiss every non-flag as confirmed health are misusing the tool. Used correctly, it improves the quality of decisions a grower was already making, not the number of decisions they have to make.

Finally, consistent image capture is non-negotiable. A team that takes careful, well-lit, close-range leaf images will see significantly better results than one that uses the system casually. The model is only as good as what it shows.

Back to the grower

The grower walking the tomato section at the start of this article is still walking it. That will not change. But what they see while walking it, and what they do with that information, can. An AI flag on a leaf that looked fine three days ago is no substitute for experience. It is a second set of eyes that never blinks.

What the Omdena project demonstrated is that computer vision for crop disease detection is no longer a research concept. It is a buildable, deployable system that can operate in real growing environments. The accuracy is not perfect, and the setup requires discipline, but the direction is clear.

The grower who starts small, learns what the system can and cannot do, and builds capture discipline into their team’s routine will be significantly better positioned than the one who waits for the technology to be fully ready. It is close enough now to be worth starting.

Somesh Utkar is a growth specialist at Omdena.