A red patch is not a diagnosis. First decide where to check on the ground.
An index is a starting point
NDVI compares near-infrared and red reflectance using (NIR − red) / (NIR + red). It is a vegetation indicator, not a standalone diagnosis or prescription. NASA explains its measurement basis; agronomic interpretation still requires context.
Start with one decision
Our proposed workflow starts with a decision: locate missing plants, confirm weeds or inspect irrigation. This determines resolution, timing, ground samples and delivery deadline. Compare full-field high-resolution capture with satellite screening followed by targeted drone inspection.
Make dates comparable
Link field ID, crop stage, acquisition time, sensor, calibration, lighting and processing version. Use comparable map scales and quality flags. Separate shadows, water and missing observations from genuine change; do not silently recolor each date.
Deliver a location and next action
The farmer-facing output should name the location, deadline, evidence photo and uncertainty. Keep raw imagery and indices available as detail. Collect ground-check outcomes to track false alarms and missed issues. Do not automatically authorize fertilizer application from an image score.
Test whether the model transfers
INIFAP’s forage-maize study recommends further validation by growth stage. Its 81% result is explained variance, not diagnostic accuracy. Fit within a research field should be distinguished from performance on a new farm.
Accept usable coverage, not attractive imagery
Our proposed acceptance process begins with the area, decision and delivery deadline. Require separate maps of captured coverage and usable analytical coverage. Check whether shadows, stitching defects, blur and positional uncertainty appear as missing data. Measure displacement from known ground points and agree tolerances separately for scouting and targeted application. Unknown boundaries must not be filled with zero values. Acceptance should include a field worker locating the indicated inspection point, rather than merely opening an image file. Agree the cost and deadline for any necessary repeat flight before commissioning.
Inspect areas without alerts too
Checking only red areas cannot reveal missed problems. Sample both alerted and non-alerted locations and inspect them at a comparable growth stage. Define the condition being assessed—disease, water stress or missing plants—before judging performance. Distinguish locations used to develop the model from those reserved for final evaluation. Record false alarms as additional scouting work and misses as delayed responses, rather than relying solely on image-level accuracy. Display whether other farms or years have been tested. When a reviewer corrects an estimate, preserve the original output and the reason.
End the analysis with an action record
An illustrative record could read: “Field A, north block; inspection required; water stress versus disease unresolved; assigned worker to inspect before the next irrigation decision.” Simplicity must not hide uncertainty. Center the farmer’s view on location, deadline, owner and outcome, with imagery and processing versions available as detail. Let the reviewer close the task as irrigation repair, continued observation or poor imagery. Exports to another service should retain field ID, observation time, units and the reason for a hold. A delivered map with no assigned follow-up should remain visible as unfinished work.