Lesson objective
Describe how an image classifier is trained on labelled pictures, and predict how a photograph taken in a shaded bed on a working property differs from the pictures the model learned from.
Plain-language explanation
Why this matters now
Pointing a phone at a leaf is the single most attractive AI feature in this trade. It takes two seconds, it returns a name, and the name arrives with a number beside it that looks like certainty. Every part of that experience is designed to feel like looking something up. None of it is.
This lesson is about the machinery: how a model learns to sort pictures, and what a real photograph on a real property does to that machinery. Module 2 gives you the measured accuracy figures and the discipline that follows from them. Take the mechanism first, because once you can picture how the model learned, the measured results stop being surprising.
A question to sit with
Think of the last plant photograph you took on a job. What was in the frame besides the plant, what was the light doing, and how much of the plant could the camera actually see?
What goes wrong
Believing the model recognises a plant. It does not know what a plant is. It learned to sort pictures into labelled groups, and it will sort your picture into one of those groups whether or not your picture belongs in any of them.
Assuming the training pictures look like your picture. Training sets are usually assembled to be useful: a clear subject, decent light, a sensible angle, often a single specimen against an uncluttered background. A photograph taken at 11 a.m. in a shaded bed, of half a leaf partly behind a downspout, with mulch and turf in the frame, is a different kind of image.
Trusting a benchmark score as a field result. A number measured on the dataset a model was built and tested on tells you how it does on that dataset. It does not tell you how it does on a customer's property.
Assuming the list of possible answers includes the right one. A classifier can only answer with a label it was trained on. If the thing in front of you is not in its labels, it will still return something.
Treating a confidence percentage as probability of being right. It is the model's internal ranking, not a measured hit rate on your kind of photograph.
The practical answer
How the training works, in four steps.
- Collect and label. Someone assembles a large set of pictures, each tagged with a label: a species, a disease, a pest, a condition. The quality and coverage of those labels sets the ceiling on everything after.
- Learn features. The model adjusts itself, over many passes, to pick up on visual patterns that separate one label from another: edge shapes, vein patterns, textures, colour distributions, the arrangement of parts.
- Test on held-back pictures. Some images are kept out of training and used to score the model. This is where the impressive percentages come from.
- Deploy. The model meets photographs nobody curated, taken by people with different phones, in different light, at different times of year, with different things in the background.
Steps 3 and 4 are where the story turns. The gap between them has a name in the research literature: performance measured on a curated dataset does not carry over to the field.
Why your photograph is harder than the training picture. Work down this list the next time an app gives you a name you are tempted to act on.
| What differs in the field | Why it matters to a classifier | What you can do about it |
|---|---|---|
| Light: deep shade, dappled shade, harsh midday sun, evening colour | Colour and texture cues shift; shadows read as marks and marks disappear | Photograph in even light; avoid dappled shade; do not use flash on a glossy leaf |
| Background: mulch, turf, gravel, fence, another plant | The model has no way to know which object in the frame you meant | Fill the frame with one subject against a plain background such as a clipboard |
| Partial view: one leaf, a branch tip, damage without the plant | Distinctive features may be entirely absent from the frame | Capture several informative views, including the whole plant and any reproductive structures |
| Season and life stage: juvenile foliage, dormancy, post-flowering | Training pictures cluster around photogenic stages | Note the date and stage, and say so when you record the identification |
| Camera and distance: motion blur, low resolution, heavy compression | Fine detail the model relies on is simply not there | Steady the phone, get closer, check the photograph before you leave the property |
| Damage and disease overlay: chewing, drought stress, herbicide effects | Confounds the visual pattern in ways the training set may not contain | Record the symptom pattern in words as well as pictures |
Six rows, six field differences. None of them is a criticism of the technology. They are simply the conditions under which you work.
A second kind of machine vision worth understanding. Not every model that sees is a classifier on a phone. Some are perception systems on machines that move. A peer-reviewed study published in Animals on 21 April 2021 by Rasmussen, Schroder, Mathiesen and colleagues tested 18 models of robotic lawn mower in collision with dead European hedgehogs, across four weight classes and three positions, scoring the results into six damage categories. The authors report that "None of the robotic lawn mowers tested was able to detect the presence of dependent juvenile hedgehogs (<200 g)", and that all models had to touch the hedgehogs to detect them. Pivoting blades, skid plates and front-wheel drive significantly increased the safety index of the mowers tested, and damage ranged from none for some models to severe injury for others.
Date that finding carefully. It is Danish work from 2021 on a specific set of machines, obstacle sensing on consumer robotic mowers has moved on since, and it is not a general claim that all machine vision misses small objects. What it does show plainly is the point of this lesson: a perception system detects what it was built and tested to detect, and something outside that set is not "low confidence", it is invisible. Module 13 covers autonomous equipment properly.
Where that leaves the app. A name from a photograph is an input to your thinking, at the same level as a hunch from across the lawn. It is not a finding, it is not a diagnosis, and it can never be the basis of a treatment decision on its own. The decision to apply a pesticide, and the product, rate, timing, site and conditions, belongs to the certified applicator, or to the certified applicator supervising a non-certified one, with the product label as the binding instruction.
Remember this
- A classifier sorts pictures into labels it was trained on; it does not recognise plants and it cannot answer outside its labels.
- Benchmark accuracy is measured on curated pictures; field conditions are different by design.
- A 2026 peer-reviewed review found controlled-dataset accuracy often decreases in real-field conditions, citing an example dropping from near 99% to 31% on external field images, in crops; transfer to turf, ornamentals and trees is untested.
- Light, background, partial views, season, camera quality and overlapping damage all move your photograph away from the training set.
- The app gives you a hypothesis. The qualified person, and the label, decide what happens next.
Landscaping example
Kestrel Tree & Shrub LLC is a fictional tree and shrub care company. Technician Bo Ferreira is at an invented property on Waverly Court at 11:40 a.m. in late June, looking at a mature maple with browning along the leaf margins on the north side of the crown. He takes one photograph: a single leaf, held up against a bright sky, in dappled shade under the canopy. The app returns a fungal leaf disease with 94% beside it. Bo has been taught what that number is, so he stops. He retakes the photograph set properly: the whole tree from the driveway, the affected branch, five leaves laid flat on the back of his clipboard in open shade, the undersides of three of them, a close view of one margin with a pen in frame for scale, and the base of the trunk. This time the app returns two candidates, neither of them the first answer, at much lower percentages. Bo also records what the app cannot see: the browning is worse on the side facing the driveway, there is a de-icing salt store at the head of that driveway, the turf between the tree and the drive is thin, and a soaker hose loops around the root plate. He writes his note as a hypothesis with a date, not a finding, and sends the photograph set and the site observations to Kestrel's licensed arborist, Priya Ambrose, who visits two days later. She rejects both app candidates on inspection and records an abiotic stress pattern consistent with the site conditions, and a soil sample goes to a lab for confirmation. No product is applied in the meantime. In the customer file the entry reads: app hypothesis, photograph set, site observations, arborist assessment, sample pending, decision maker named. Had Bo acted on the first 94%, Kestrel would have billed a fictional customer for a fungicide application on a tree that did not have a fungal problem, and the salt store would still be there.
Responsible-use warning
Never let a percentage beside a species or disease name stand in for a diagnosis. The model sorted your picture into the closest label it was trained on, whether or not the right answer was among those labels, and a benchmark accuracy measured on curated images says nothing about a photograph taken in dappled shade against the sky. Retake the picture set properly, record what the camera cannot see, and route the question to the qualified person. No product is chosen, no rate set and no timing decided from an app result. The label on the container is the binding instruction and the certified applicator owns the decision.
Human-review requirement
The qualified person confirms or rejects every identification before it affects a treatment, a bid or a written report to a customer. Where the next step is a pesticide application, that person is the certified applicator, or the certified applicator supervising a non-certified applicator, who owns the product, rate, timing, site and conditions with the label as the binding instruction. Where the subject is a tree's condition, it is the arborist or, in a state that licenses tree work, the licensed tree expert. The reviewer checks that the photograph set and site observations were recorded, that the app output was logged as a hypothesis with its date, and that no crop accuracy figure was quoted as if it applied to turf, ornamentals or trees.
Practical exercise
Test the mechanism on your own photographs. Use fictional or de-identified property references in anything you write up.
- Choose one plant on a property you can visit, or one specimen you can bring to the shop. Do not choose anything a treatment decision currently depends on.
- Take a single deliberately poor photograph: one partial leaf, in dappled shade, with a busy background. Record the app result and the number beside it.
- Take a proper set: whole plant, affected part, upper and lower leaf surfaces, any flower or seed structure, and one frame with an object for scale, all in even light against a plain background.
- Run the set and record every result and number. Note whether the top answer changed between step 2 and step 3.
- Write down three things about the plant or the site that no photograph in your set can show, such as irrigation history, recent applications, or what is uphill of the root zone.
- Using the six-row table in this lesson, name which field differences applied to your poor photograph and which you fixed.
- Write the identification as a hypothesis: what it might be, what would confirm it, what would rule it out, and who decides.
- Name the qualified person in your company who would own the next step if this were a live job, and state why it is that person and not you if you are not that person.
- File the photograph set, your notes and the app outputs together, and note the date and time of day.
Knowledge check
-
What is an image classifier actually doing when it names a plant in your photograph?
Answer guidance: It is sorting the picture into one of the labels it was trained on, based on visual patterns it learned from labelled training pictures. It does not recognise plants, and if the correct answer is not among its labels it will still return one of them.
-
Why does a high benchmark accuracy not predict performance on your property?
Answer guidance: Benchmark figures are measured on held-back pictures from a curated dataset. The 2026 Plant Pathology Journal review found that effectiveness often decreases under real-field conditions due to environmental variability, limited training data and deployment constraints, citing an example that dropped from near 99% to 31% on external field images, in crops.
-
Name four ways a working photograph differs from a training picture.
Answer guidance: Any four of: light such as deep or dappled shade and harsh sun; a cluttered background with mulch, turf or structures; a partial view showing only one leaf or a branch tip; season and life stage; camera distance, blur and compression; and damage or stress patterns overlaying the plant.
-
What did the 2021 robotic mower study find, and what must you say when you cite it?
Answer guidance: Testing 18 mower models against dead European hedgehogs in four weight classes and three positions, no mower detected dependent juveniles under 200 g and all had to touch them to detect them; pivoting blades, skid plates and front-wheel drive raised the safety index. Cite it as Danish 2021 work on specific machines, note that consumer obstacle sensing has moved on since, and do not treat it as a general claim about all machine vision.
Sources and further reading
- The Plant Pathology Journal (Korean Society of Plant Pathology), peer-reviewed — Artificial Intelligence-Driven Plant Disease Detection and Diagnosis: A Comprehensive Review of Deep Learning Approaches, Multimodal Sensing Technologies, and Future Perspectives in Precision Agriculture (2026-04-01). Accessed 2026-09-15.
- Animals (MDPI), peer-reviewed; authors Rasmussen SL, Schrøder AE, Mathiesen R, et al. — Wildlife Conservation at a Garden Level: The Effect of Robotic Lawn Mowers on European Hedgehogs (Erinaceus europaeus) (2021-04-21). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.