AI for Farmers, Gardeners, and Food Systems

Use local records to compare plans, and make every missing measurement visible.

A gardener uploads a photograph of a struggling plant and receives a confident diagnosis. A farm manager asks for an irrigation schedule and receives a neat weekly table. Both answers may overlook the most important facts: what is happening in the soil, how much water the system actually delivers, and whether the symptoms have more than one possible cause.

AI can help organize observations and compare scenarios. Its usefulness increases when the task is tied to local records and a specific decision. A good planning worksheet tells you what the calculations imply, what assumptions they depend on, and what you must check before acting.

This article uses a fictional irrigation example. The numbers teach unit handling and scenario comparison. They are not a watering recommendation for a particular crop or site.

Separate planning from identification and control

Several different activities can hide inside the phrase “AI for agriculture.” A planning assistant might organize field logs. An image system might suggest possible causes of leaf damage. A forecasting model might estimate demand or yield. A control system might operate irrigation equipment.

Those tasks require different evidence. A text assistant that summarizes a field log has not demonstrated that it can diagnose disease. A model that forecasts water demand has not demonstrated that it can safely operate a pump.

Choose a narrow question first: “Given a supplied net water deficit, measured area, assumed application efficiency, and measured flow rate, calculate the corresponding gross volume and runtime.” This question can be answered transparently. Determining the correct deficit is a separate agronomic task.

For food systems, the same approach works beyond fields. AI can organize harvest records, compare demand scenarios, or summarize storage observations. Keep food-safety decisions and operating limits tied to the applicable verified process rather than asking a general model to invent them.

Assemble a local data packet

A useful irrigation packet might include the growing area, crop and growth stage, soil observations, root-zone assumptions, rainfall records, water source, system configuration, measured flow, and the time of each observation.

Dates matter. Yesterday’s rain can change today’s decision. A flow measurement from before a system modification may no longer apply. A soil reading from one location may not represent the whole area.

Separate the entries into measured, externally sourced, assumed, and unknown. “Application efficiency: 80%, assumed for this exercise” is much more informative than an unlabeled number in a spreadsheet.

Include how measurements were collected. A soil reading needs a location and depth, not merely a value. A flow measurement needs the operating configuration. An area estimate should identify whether paths and non-irrigated sections were excluded.

Soil properties also affect the interpretation of available water. Extension materials can explain concepts such as field capacity and rooting depth, but a guide written for one setting is not automatically a prescription for another. For example, New Mexico State University’s turfgrass irrigation guide provides a setting-specific discussion; its turf recommendations should not be transplanted into a vegetable plan without appropriate review.

Work through the units before producing a schedule

Our fictional teaching packet contains:

A depth of one millimeter over one square meter corresponds to one liter. You can derive that relationship: 0.001 meters multiplied by one square meter is 0.001 cubic meters, or one liter.

The net volume is therefore:

100 m² × 5 mm × 1 L/(m²·mm) = 500 L.

Under the exercise’s definition, application efficiency is the fraction of gross applied volume that supplies the intended net amount. Gross volume is:

500 L ÷ 0.80 = 625 L.

At the supplied discharge, the corresponding runtime is:

625 L ÷ 250 L/hour = 2.5 hours.

Every line carries its units. That makes a common mistake easier to catch: multiplying by 0.80 would produce only 400 liters, which cannot deliver a 500-liter net requirement under this assumption.

The runtime is a calculated consequence of the inputs. It is not yet an operating instruction. The calculation does not establish whether continuous application is suitable, whether the soil can accept the water at that rate, or whether rain has already reduced the deficit.

Compare a second scenario fairly

Suppose a second hypothetical configuration has an application efficiency of 0.90 and the same measured discharge. Keeping the assumed net requirement fixed gives:

The difference is approximately 69.4 liters per application. It is a scenario result, not evidence that installing a particular product will produce that improvement.

For an economic illustration, assume water costs 0.004 fictional currency units per liter. The corresponding water costs are 2.50 and approximately 2.22 units. The difference is about 0.28 units per application. These numbers exclude equipment, labor, pumping energy, maintenance, and any effect on crop performance.

This is a useful place to challenge the AI summary. “Upgrade saves money” goes beyond the calculation. “Under the stated assumptions, the second scenario uses less gross water and has a lower water-only cost” accurately describes the result.

Understand what evapotranspiration can and cannot supply

A more advanced planning workflow may use reference evapotranspiration and crop coefficients. FAO describes a standard approach in which crop evapotranspiration is calculated as a crop coefficient multiplied by reference evapotranspiration. The standard conditions and the appropriate coefficient matter; the relationship is not a license to invent local values. See FAO’s crop evapotranspiration guidance.

Ask AI to show where each coefficient came from, which crop stage it represents, and whether the source’s conditions match the planning question. If the model supplies a coefficient without an inspected source, move it to the “unverified” column.

Even a suitable evapotranspiration estimate does not answer every irrigation question. Effective rainfall, existing soil water, root-zone capacity, system behavior, and local restrictions may change the action. Organize those inputs explicitly rather than burying them in a disclaimer after a confident schedule.

A useful output might be: “The worksheet can calculate gross volume once the net deficit is confirmed. It cannot yet determine a field schedule because effective rainfall and application limits are unresolved.” That is a productive result: it tells the operator what information would unlock the next step.

Treat images as observations with limits

A photograph can document visible symptoms, distribution, or change over time. It often cannot distinguish all possible causes. Ask AI to describe what is visible before proposing explanations.

A field record might say: “Three lower leaves show yellow areas in the supplied image; the image does not establish soil moisture, root condition, or pathogen identity.” The next task is to identify what observation or qualified consultation would distinguish the plausible explanations.

Preserve the original image, capture date, location, and relevant growing conditions. Avoid labeling the image with a final diagnosis before it has been verified; that label can later contaminate a training dataset or mislead another reviewer.

For unfamiliar or consequential crop problems, an appropriate extension service or qualified specialist can help determine the necessary evidence. The role of AI is to make the packet clearer and the questions more specific.

Turn a plan into a monitored trial

A small trial should have a defined question, recorded starting conditions, a comparison where appropriate, and a stopping or review rule chosen by the responsible person. Document what will be measured and when.

For the irrigation worksheet, useful follow-up records could include actual metered volume, observed application problems, rainfall, and relevant soil observations. Compare actual volume with the calculated volume before claiming that the plan performed as expected.

Do not attribute a crop change to the irrigation adjustment simply because it happened afterward. Weather, pests, growth stage, and other management changes may also have changed. AI can help maintain a change log so the later interpretation has a fairer account of competing explanations.

A reusable prompt

Compare these agricultural planning scenarios using only the supplied local measurements and explicitly labeled assumptions. Show units in every calculation. Separate measured, sourced, assumed, and missing inputs. Identify which agronomic judgments, operating limits, and expert checks must be resolved before action. Do not diagnose a plant from an image alone or turn a scenario calculation into an irrigation instruction. Propose a field log that would let us compare the plan with observations.

For students: learn dimensional reasoning with a garden model

Use a fictional bed, a school-approved garden exercise, or a paper dataset. Calculate the volume represented by several depths over different areas. Explain why doubling area doubles volume when depth stays constant.

Students in biology can focus on the observations needed to interpret plant response. Mathematics students can examine sensitivity to efficiency assumptions. Business students can distinguish water-only savings from total costs.

If a school garden is involved, follow the teacher’s operating instructions. You can learn the calculation without changing the watering system. Your report should identify which numbers you measured and which were provided for the exercise.

Practice: complete the irrigation worksheet

Recalculate the example for an area of 80 m² with the same assumed 5 mm deficit, 80% efficiency, and 250 L/hour discharge. Then calculate the 90% efficiency alternative. Show net volume, gross volume, runtime, and the difference between the alternatives.

List at least five pieces of information you would need before adopting either result as a real schedule. Include a timing question and a question about the site’s ability to receive the water.

Completion check: Your first scenario gives 400 L net, 500 L gross, and 2 hours. Your second gives approximately 444.4 L gross and 1.78 hours. Every input has a status, every calculation has units, and no unmeasured local condition is presented as established.

Stretch: Design a two-week field log with separate columns for planned volume, actual volume, rain, observations, and changes in management. Explain what the log could reveal and which causal claims it would still not support.