AI Crop-Quality Detection & Workforce Allocation

The problem: quality and plant-protection decisions made by guesswork

A large tea-estate group — 3,200 hectares spread across seven estates, with more than 10,000 field workers — came to us with a problem that is common to almost every large agricultural operation, even if the crop changes. The value of the harvest depends on a quality signal that is hard to see at scale, and the cost of protecting that harvest depends on knowing exactly where the threats are. At their scale, both decisions were being made the same way they had been for decades: through manual spot-checks, supervisor experience, and a fair amount of educated guessing.

For tea specifically, the prize is the two-leaves-and-a-bud growth — the tender top composition that commands the highest price and defines a premium pluck. Spotting where that growth is concentrated, across thousands of hectares, is the difference between a good harvest and a mediocre one. But a supervisor can only inspect so many bushes in a day, and human judgement drifts. Two inspectors will grade the same field differently; the same inspector will grade differently before and after lunch. The result is that harvesting effort gets spread evenly instead of being concentrated where the most valuable leaf actually is.

The plant-protection side of the problem was worse. Pests and disease do not announce themselves politely. By the time an outbreak is obvious to the eye, it has usually already spread. So the estate did what most operations do when they cannot see the threat precisely: they treated everything, everywhere, on a calendar. Chemicals and crews were applied as a blanket across all seven estates, which is expensive, environmentally heavy, and — paradoxically — still misses the early-stage problems that matter most. Spending was high and effort was thin at exactly the points where it should have been concentrated.

How the system works

We built a computer-vision system that does two jobs at once: it grades leaf quality, and it flags potential pest and disease issues early — then turns both signals into a map that tells the operation where to send people and resources. At the core sits a convolutional neural network based on the VGG-19 architecture, trained to recognise the visual composition of the high-value pluck and the early visual markers of plant stress.

VGG-19 is a deep, well-understood image classifier, and that maturity matters in a field setting. It is reliable, it is explainable enough that agronomists can trust what it is responding to, and it performs well on the kind of fine-grained texture and colour distinctions that separate a premium two-leaves-and-a-bud pluck from coarser growth. The pipeline runs in three clear stages:

• Detect. The VGG-19 model grades leaf composition from field imagery, identifying the prized two-leaves-and-a-bud growth, and simultaneously spots the early visual signs of pest damage and disease before they become a visible outbreak.
• Map. Those detections are aggregated into a spatial picture of the estates: where the high-quality growth is concentrated, and where at-risk areas are emerging. Quality and risk stop being anecdotes and become coordinates.
• Allocate. Crews and plant-protection effort follow the map rather than the calendar. Harvesting teams are directed to the most valuable growth; chemicals and labour are applied to the areas that actually need them.

The shift is subtle but enormous. The estate moves from blanket treatment and even effort to targeted treatment and concentrated effort — the same principle that makes precision agriculture work everywhere it is applied. You stop spending on the 80% of the estate that is fine, and you put that budget and that labour into the 20% that determines the outcome.

Data and approach

A model is only as good as the data behind it, and agricultural data is messy. Field imagery varies with light, weather, season, and the angle of capture; the same bush looks different at dawn and at noon, in sun and under cloud. Our machine-learning work here was as much about disciplined data preparation as it was about the model itself. We worked with the estate's own imagery so the system learned their cultivars, their conditions, and their definition of a premium pluck — not a generic benchmark from a different crop in a different climate.

That is a principle we apply to every vision project: a model trained on someone else's tea, or on lab conditions, will not survive contact with a real estate. Transfer learning from a strong base like VGG-19 gives the model a head start on general visual features, and fine-tuning on the client's labelled examples teaches it the specific distinctions that matter to their grading standard. Agronomists stayed in the loop throughout, because the goal was never to replace expert judgement — it was to scale it across 3,200 hectares that no team of supervisors could ever cover by eye.

Crucially, the output is decision support, not autopilot. The system tells the operation where quality is concentrated and where risk is emerging; experienced managers decide how to act on it. That keeps the people who know the land in control, and it keeps the system honest about its own uncertainty.

Results

The outcomes were measured on the client's own estates, and they were substantial. By mapping at-risk areas precisely and treating only where treatment was warranted, the operation cut its plant-protection spend dramatically:

• Annual plant-protection cost fell from €800K to €225K — a reduction of more than 70%, driven by replacing blanket treatment with targeted application.
• The quality of tea leaves harvested improved by 14%, because harvesting effort was concentrated on the high-value two-leaves-and-a-bud growth the model identified and mapped.

Those two numbers reinforce each other. Lower input cost and higher output quality at the same time is the signature of a precision approach working as intended: you are not trading one against the other, you are removing waste from both sides. The plant-protection saving alone pays for the system many times over; the quality lift compounds the return year after year.

Who it's for, and the ROI

This is not really a story about tea. It is a story about any operation where quality is visible but hard to grade at scale, and where protecting an asset means knowing precisely where the threat is. The same pattern — vision to grade, mapping to locate, allocation to act — applies far beyond a tea estate. We have built closely related systems for industrial inspection, including automated concrete crack detection on production lines and predictive maintenance for industrial plants, and the same computer-vision foundations support quality and defect work across manufacturing and asset-heavy operations in logistics.

The ROI logic is straightforward and it transfers cleanly. Wherever you are currently spending on a blanket basis — uniform treatment, uniform inspection, uniform effort — there is almost always a precision version that costs less and delivers more. The two questions worth asking are: how much are we spending evenly across an area where the problem is concentrated? and how much value are we leaving on the table by not being able to see quality at scale? If either number is large, vision-led targeting tends to pay back fast. You can see indicative engagement tiers on our pricing page.

How we'd run a pilot

We never ask a client to take results like these on faith. The honest way to start is a focused pilot on your own data, with success defined before we begin. For an operation like this, that means: agree the quality definition and the risk signals that matter to you; gather a representative set of your own field imagery across conditions; fine-tune the model on your labelled examples; and run it against a manual baseline so the comparison is real, not theoretical.

The pilot ends with a clear go / no-go and verified numbers from your estate, not a benchmark from someone else's. If it works — and on a well-scoped problem like quality grading and early detection, it usually does — we move to a full AI implementation that runs in production, integrated into the way your teams already plan and deploy crews. If you have a problem that looks like this one, the fastest way to find out what it is worth is to tell us what you are trying to fix, and we will scope a pilot that reports real numbers on your own ground.