The drone sweeps the field in a lawnmower pattern. Under each pass it samples two synthetic reflectance bands per plant (near-infrared and red) and turns them into the same vegetation index real crop-monitoring drones use:
NDVI = (NIR − Red) / (NIR + Red) range: −1 .. +1
classify(NDVI):
NDVI < sensitivity → diseased (red)
sensitivity ≤ NDVI < sensitivity+0.18 → stressed (yellow)
NDVI ≥ sensitivity+0.18 → healthy (green)
measured_NDVI = true_NDVI + noise · (altitude / 15m)
- Disease sensitivity — the classifier's decision threshold. This is the precision/recall dial every real detection model ships with: push it down and the AI flags borderline plants too, catching more true disease at the cost of healthy plants wrongly flagged.
- Altitude — higher flights cover ground faster but each pixel averages a larger patch of canopy, so sensor noise grows and small disease pockets get blurred into their healthy neighbours — the same resolution/coverage trade every multispectral drone survey makes.
- Flight speed — how fast the drone completes one full field pass; coverage % and the plant colours update live as it flies over each row.
Forecast yield loss is a simple regression from mean field NDVI — the same kind of index-to-yield model agronomists calibrate against historical harvest data to turn a health map into a bushels-per-acre estimate before the crop is even close to harvest.