Diagnosing a crop problem from a photo, and knowing when to escalate
The situation
Something is wrong in the field and the question is whether it matters today. Historically that meant waiting for a farm advisor for problems that turned out to be nothing, or ignoring things that turned out to be serious. A photographed symptom plus a conversation gets you a short list of possibilities fast. What makes it responsible practice rather than guessing is a written rule, decided in advance, about what you do not handle yourself.
Steps
-
Write the escalation rule first, while nothing is wrong
A card in the truck, or a note on your phone
Before any of this is used in the field, decide and write down what always goes to a professional regardless of what the model says: anything spreading between visits, anything on more than a set share of plants, anything resembling a quarantine pest, anything where treatment would be a pesticide application.
What you only learn by doing it: Written in advance, this survives the moment when you want the answer to be the easy one. Written afterwards, it is a rationalisation.
-
Photograph it properly, which takes four shots not one
Your phone camera
The whole plant, the affected part close up, the underside of a leaf, and the pattern across the planting. Include something for scale. Note what is adjacent.
What you only learn by doing it: The distribution shot is the one people skip and the one that carries the most information. Scattered single plants, an edge effect and a spreading patch mean entirely different things, and no close-up can tell them apart.
-
Describe conditions with the photos, then ask for a differential
Google (Gemini Pro) OpenAI for Education
Crop and variety, growth stage, recent weather, irrigation, what was applied and when, whether it is spreading. Then ask explicitly for several possibilities ranked with reasons and the observation that would separate them — not for an identification.
What you only learn by doing it: Ask what it would need to see to rule each one out. That question converts a confident guess into a plan, and it is also the fastest way to find out the photos were not good enough.
-
Apply the rule you wrote in step one
Your local extension or UC Cooperative Extension farm advisor
Compare what came back against the escalation rule and follow the rule, not the model. If it trips, the next contact is your farm advisor, extension office or county agricultural commissioner.
What you only learn by doing it: For students, this is the step being assessed. Correct identification is luck; escalating correctly is competence, and it is what an employer is actually trusting them with.
-
Log it, then close the loop on what it actually was
A shared folder or spreadsheet
Record the photos, the differential, what you did and the date. Then record the real answer when you learn it. A log of guesses teaches nothing; a log of guesses against outcomes is how calibration is built.
What you only learn by doing it: Over a season this becomes the most valuable thing produced here, and the only part that makes next year’s judgement better rather than just faster.
Where this breaks down
Confident misidentification is the normal failure, not the rare one. Many diseases, deficiencies and herbicide injuries look alike in a photograph, and the model will name one without signalling how thin the evidence is. Treat any single answer as a hypothesis. A differential with reasons is useful; a confident name is a trap.
The costs are asymmetric and that should drive the whole workflow. Missing a quarantine pest or a fast-moving disease can cost a crop or a region. An unnecessary spray costs money, beneficial insects and sometimes a residue violation. Neither error is cheap, so the escalation rule has to be written before you are standing in the field wanting an answer.
The pesticide label is the law. A model’s treatment suggestion is not a label, carries no legal weight, and may name a product not registered for your crop, your state or your situation. In California, apply nothing without the label and your county agricultural commissioner’s requirements in hand.
Some pests and diseases are reportable. A suspected quarantine organism goes to the county agricultural commissioner or the state department of agriculture immediately, not into a chat log.
Provenance: adapted from Hiroki Tomiyasu, who farms roughly 100 hectares in Hokkaido, interviewed in OpenAI’s ChatGPT for Pros newsletter, 1 June 2026. Tomiyasu photographs abnormalities on the spot and describes the practical value as triage: guidance “ranging from minor symptoms that do not require consulting a specialist to diseases that need urgent intervention.” The sequence, the checkpoints and the cautions below are our construction, written for a classroom. Treat the source as one practitioner’s documented experience in a vendor publication, not as evidence that this works generally.