- Only 27 percent of US farms and ranches used any precision agriculture practice in the year to June 2023, according to the government review of USDA reporting.
- Field accuracy depends on growth stage. One study across 11 weeks of weed growth found detection improving as plants matured, which means seedling stage is the hard case and the one that matters for spraying.
- The best reported figures come from curated image sets, not from a moving implement. A segmentation model cited in the 2026 review scored 79.1 percent on edge hardware against 85.4 percent on a full GPU.
- The binding constraint is the computer on the machine. Inference has to run in the field with no connectivity, so you get a lightweight model rather than the best one.
- Ask what crops and growth stages it was trained on, what it does with a species it has never seen, the frame rate on the hardware being sold, and who owns the images it collects.
The demo was filmed on a bright morning in a clean trial plot, rows straight, weeds obligingly distinct from the crop. The camera found every one. Six months later the same rig is working a field in late June, canopy closing, a week after rain, with volunteer wheat coming through that looks exactly like what it is supposed to leave alone. That gap is the whole subject, and almost nothing written for buyers addresses it.
AI in agriculture is one of the places where the claims are most testable, because the machine either kills the weed or it does not. It is also where the published evidence is thinnest in the direction buyers need. Research reports precision and recall on image datasets. A farm needs to know what happens across a season.
How many farms are actually using this?
Fewer than the coverage suggests. The US Government Accountability Office reviewed the question in its January 2024 report Precision Agriculture: Benefits and Challenges for Technology Adoption and Use and found that only 27 percent of US farms or ranches used precision agriculture practices to manage crops or livestock, on 2023 USDA reporting covering June 2022 to June 2023.
The report makes a point worth holding onto: variable rate fertiliser application and yield monitoring have been available since the 1990s, and adoption is still low. The barriers it names are not about model accuracy at all. High up front acquisition costs put the newest equipment out of reach for farms without access to capital. Disputes over who controls farm data slow wider use. Absent uniform standards, devices from different suppliers often do not work together.
If you are evaluating a system, that last one deserves early attention. A camera system that cannot write to the controller you already run, or whose data you cannot export, is a second purchase pretending to be one.
Why does crop vision fail outdoors when it works indoors?
Because almost everything the model relies on moves. A conveyor in a packhouse has fixed light, a fixed background and a fixed distance to the object. A field has none of those, and the 2026 review of the weed detection literature is specific about which variables do the damage.
That review, Deep learning based approaches for weed detection in crops, published on 9 January 2026, states that object detection performance is highly affected by variations in illumination, crop growth stages, weed density and occlusion by crops. Segmentation accuracy suffers from shadows and from colour similarity between weed and crop. Classification models generalise poorly across lighting conditions. Small weeds get missed, and accuracy drops in dense canopies.
Each of those has a date attached in practice. Shifting light is a time of day problem, worst at dawn and dusk when operators want to extend the working window. Occlusion is a calendar problem: before the canopy closes the camera sees soil and leaf, after it closes the weed is behind the crop. Growth stage is the subtle one, because a plant at two leaves and the same plant at flowering are different objects to a detector, and the economically important moment is the early one.
The evidence on that last point is unusually good. A 2025 study using object detection models across 16 weed species at 11 weekly growth stages, built on a dataset of 203,567 labelled images, reported mean average precision of 0.904 on its test set for the stronger of the two models tested, against 0.840 for the other. The sentence that matters for a buyer is quieter: both models showed improved accuracy as plants matured. Detection is easiest when the weed is big, and spraying is most useful when it is small.
One more thing the published work keeps returning to, and buyers consistently underrate. Precision and recall reported on a curated image set are not a forecast of performance from a moving implement. A test set is drawn from the same collection as the training images, which means it shares their light, their soil and their season. The field does not. Treat any single headline percentage as an upper bound measured under conditions you will not meet, and ask for the figure under conditions you will.
Which tasks are mature enough to buy?
Four jobs get sold as one capability and they are at very different stages. This table sorts them by what each one demands of the system rather than by how good the demo looks.
| Task | What it demands | Where it currently stands | What failure looks like on your farm |
|---|---|---|---|
| Weed identification for targeted spraying | Species level calls at seedling stage, in real time, on a moving boom | Workable on known species in known crops; weakest exactly at the early stage that matters | Missed weeds that set seed, or crop damage from a false hit |
| Crop and disease scouting | Recognising symptoms that vary by cultivar and weather, often from a drone pass | Useful as a triage layer that tells you where to walk | Confident labels on the wrong disease, which sends the wrong spray |
| Yield estimation and counting | Counting fruit or heads reliably through partial occlusion | Reasonable at field scale, poor at plant scale once the canopy closes | A forecast that drifts, and marketing decisions made against it |
| Harvest readiness | Judging maturity from colour and shape under changing light | Earliest stage of the four; highly crop specific | Picking a block early or late, which is a quality problem not a yield one |
The pattern across the four is consistent. Anything that asks the system to tell you where to look is further along than anything that asks it to act without you. A scouting pass that flags six zones to inspect is useful even when two flags are wrong. A sprayer deciding what to kill at 12 kilometres an hour has no such tolerance, which is the same distinction we drew between vision jobs in what the different computer vision tasks actually do.
Why is the dataset the real bottleneck?
Because the model is only as good as the labelled pictures behind it, and labelling pictures of weeds is slow, skilled work. The 2026 review puts this plainly: the main bottleneck is not network architecture but the substantial cost and difficulty of building high quality annotated datasets. Architectures are published and shared. An annotated dataset of fifty thousand correctly identified seedlings in your soil type is not.
Two consequences follow for a buyer. The first is species coverage. The review notes that weed diversity and long tail distributions leave generalisation to rare or unseen species weak, which is a technical way of saying that the common weeds in the vendor home market are well covered and yours may not be. The second is that scale shows. The growth stage study needed 203,567 labelled images to cover 16 species across 11 weeks, and that was for one controlled research question. Any vendor claiming broad coverage across many crops and regions is claiming a labelling effort of a size worth asking about.
This is also why the first season on a new farm often underperforms the sales material and then improves. You are contributing the data that makes the model work in your conditions. That is a reasonable arrangement if everyone says so out loud and the contract reflects it, and a bad one if you are paying full price for the privilege and the resulting model follows the vendor to your neighbour.
What is the constraint nobody mentions in the brochure?
The computer bolted to the implement. Inference has to happen where the sprayer is, at the speed the sprayer moves, usually with no usable connectivity. So the model is not chosen for accuracy. It is chosen to fit the hardware.
The 2026 review is blunt about the tradeoff. Segmentation, which is the most spatially precise approach and the one that handles occlusion best, carries a computational load that makes it challenging for real time deployment on edge devices or drones. Lightweight detectors such as YOLOX tiny and MobileNet SSD can run on embedded hardware, and they may sacrifice accuracy to do so. The review cites one segmentation study scoring 79.1 percent bounding box average precision on an AGX Xavier edge device against 85.4 percent on a full GPU.
Six points of accuracy for the privilege of running in a field. That is the honest exchange rate, and it is why a vendor demo on a laptop proves very little about what the rig does at working speed. The weed growth stage study reported its faster model running at 7.28 frames per second, which it described as making it more suitable for real time use. Work out what that means at your operating speed before anybody talks about percentages.
This is a general pattern rather than an agricultural curiosity, and the construction version in what a site camera can really see fails along the same lines. Any system that has to decide on a machine, away from a data centre, pays the same tax, and we traced the equivalent gap between laboratory robotics results and working deployments in why robot demos do not survive contact with a real site.
Drone or machine mounted, and does it matter?
It decides which of the four tasks you can realistically do. A drone pass gives you area coverage at a moment in time, which suits scouting and yield estimation, where a map of where to look is the deliverable. Nothing is acting on the result at the moment of capture, so inference can happen afterwards on a decent computer, and the hardware limit relaxes.
A camera on an implement is the opposite. The decision has to be made as the nozzle passes, so the model runs on the embedded box at working speed, and that is where the accuracy discount described above applies. Anything sold as real time intervention sits on this side, and should be judged against the frame rate it achieves rather than against the benchmark numbers in a paper.
There is a middle case worth knowing about, which is mapping on one pass and acting on a later one. Fly the field, build a prescription, apply it with a machine following the map. Accuracy requirements drop because nothing is being decided at speed, and the cost of being wrong is bounded by the resolution of the map rather than by a single frame. For many operations this is the version that pays first, and it is rarely the version being demonstrated.
What should you ask before signing?
Six questions. The useful property of all of them is that an honest answer is short and a weak answer is long.
- Which crops and growth stages was the model trained on? If your crop or your weed spectrum is not in the training data, you are the training data. Ask what the first season costs you.
- What happens when it meets a species it has never seen? There are two designs: treat the unknown as crop and leave it, or treat it as weed and hit it. Both are defensible, and you need to know which one you bought.
- What is the frame rate on the hardware you are selling me, at my working speed? Not on a workstation. On the box that ships, in the field, with the camera count you are quoting.
- Does accuracy get reported back, and can I see the misses? A system that logs what it skipped is auditable. One that only reports what it hit is a marketing instrument.
- Who owns the imagery and the agronomic data it collects? The government review names data control as a live barrier. Data ownership belongs in the contract, including what happens to the imagery if you leave.
- What does retraining cost when I change crop or region? This is the question that turns a capital purchase into a subscription, and it is rarely volunteered.
Does the money work?
It depends on one number, and it is not an accuracy figure. It is how much you currently spend on the thing the system replaces, per hectare, per season.
The return comes from two places. Herbicide reduction, where a sprayer that treats only weeds uses a fraction of a broadcast rate, and labour, where scouting or hand weeding hours fall. Both are real and both are specific to your operation. A high value row crop with expensive selective chemistry and a tight labour market gets a different answer from a cereal block where the broadcast rate is cheap.
Against that sits equipment cost, which the government review flags as the main barrier for farms without capital access, plus the retraining cost when you rotate into a crop the model does not know, plus the hours somebody spends keeping cameras clean and calibrated. The last one is routinely left out of the business case and is not small.
The sane test is a single block for a single season, with the comparison recorded by hand. Spray rate per hectare before and after, weed escapes counted at the same growth stage, hours logged. One honest block beats any whole farm projection, and the arithmetic is the same discipline we applied to working out what an AI tool really costs a small operation. If you also sell direct, the same measuring habit applies to the shopfront: our guide to building the online shop is the other half of a farm business that has stopped guessing.
Who should buy now, and who should wait?
The dividing line is not farm size in acres, it is how concentrated your spend is on a single operation that a camera could narrow. If one job, usually herbicide on one crop, accounts for a large share of variable cost on a block you farm the same way every year, the measurement is worth doing now. If your cost is spread across many crops in small areas, the retraining question lands on you repeatedly and the economics get much harder.
That matches what the government review found about adoption patterns, where it pointed to smaller holdings and diversified cropping as plausible reasons for lower uptake in some regions. A diversified vegetable operation changes crop, spacing and weed spectrum several times a season. Each change is a new recognition problem. The same camera that pays for itself on a single cereal or cotton block becomes a research project.
Waiting is a legitimate decision and it has a cost of its own, which is that the data your neighbours are contributing is improving the models you will eventually buy. Neither choice is obviously right. What is clearly wrong is buying on a demo, which is the one path that guarantees you find all of this out at full price in your own field.
If you do wait, the useful thing to do meanwhile costs nothing. Record, by block, what you actually spend on the job you would be automating, and count weed escapes at a fixed growth stage each season. Two seasons of that gives you a baseline no vendor can argue with, and it is the only way to tell afterwards whether the purchase worked. The same before and after discipline is what makes any forecasting claim checkable, which we went through in how to choose a forecasting method you can actually verify.
What is likely to change next?
Three things, with different confidence levels. Edge hardware keeps getting faster, which shrinks the six point gap between field and workstation performance. Annotation remains the bottleneck the literature names most often, so the methods that reduce labelling cost are the ones to watch. And species coverage is widening, which matters most to anyone outside the big commodity crops, who currently gets a model trained for somebody else's field.
What is not about to change is the physics. Light will keep moving, the canopy will keep closing, and seedlings will keep looking alike. A buyer who understands those three facts will read any demo correctly, whatever the brochure number says.