Early disease detection

On this page 11

See the infection before the eye can

Multispectral imaging picks up stress, fungal infection and nutrient shortfall days before symptoms are visible from the cab.

CategoryDetect
Slugplant-disease-detection
CadenceWeekly through the main growth stages, tightening around high risk weather
PayloadFive band multispectral, downwelling light sensor, RGB reference camera
Needs a visit firstNo

The problem

By the time a farmer sees a disease from the tractor seat, the infection has usually been established for a week or more and the treatment window has narrowed to a curative spray instead of a protective one. Walking the field catches it earlier, but nobody has the hours to walk every hectare weekly.

How it works

StepWhat happens
CaptureA multispectral camera records red edge and near infrared bands alongside visible light
IndexThe bands are combined into vegetation indices that expose chlorophyll loss and cell structure damage
CompareEach flight is compared against the field's own history, so a wet spring baseline does not read as disease
FlagDivergent patches are ranked by severity and area, with a scouting route so you can ground truth the worst first

The third step is the one that separates a working system from a demo. An absolute NDVI threshold produces a map of soil type. A comparison against the same field's own previous flight produces a map of what changed.

What you get

  • Index maps per flight
  • Change layer against the previous flight
  • Ranked scouting list
  • Ground truth capture form

What the dashboard measures

  • Affected area per flight
  • Change since the previous flight
  • Severity distribution
  • Days from flag to ground truth

Where it matters most

Winter wheat, potatoes, vineyards, orchards, hops.

Build it

What "multispectral" has to mean

Five bands, radiometrically calibrated, with a downwelling light sensor. Anything less does not survive contact with a season.

BandCentreUsed for
Blue~475 nmReference, water, some pigment work
Green~560 nmChlorophyll reference
Red~668 nmChlorophyll absorption
Red edge~717 nmThe band that carries early stress. This is the one you cannot do without
Near infrared~840 nmCell structure and biomass

Indices worth computing: NDVI (saturates early, still expected by everyone), NDRE (the workhorse for a closed canopy), CIred-edge, and MCARI/OSAVI for soil background suppression.

Calibration is not optional

Two flights a week apart under different cloud are not comparable unless you correct for illumination. That means:

  1. A downwelling light sensor on top of the aircraft, logging incident irradiance per band, per frame.
  2. A reflectance panel imaged on the ground before and after every flight.
  3. Vignetting and lens correction per band, from the camera's calibration data.

Skip these and your "change layer" is mostly a record of the weather.

Hardware

ItemWhyEU sourceUS sourceIndicative
Multispectral camera, calibratedThe whole capabilityMicaSense RedEdge-P via EU dealersAgEagle, Advexure€9,000 to €18,000
AlternativeSix band, similar classSentera 6XSentera€8,000 to €15,000
Integrated aircraftCheapest route to a calibrated five band dataset if you are not buildingDJI Mavic 3 Multispectral via SolectricAdvexure€4,500 to €7,000
Budget multispectralThree separate single-band cameras, hand calibratedMAPIR Survey3W OCN and NIRMAPIR€400 to €700 each
Downwelling light sensorComparable flights across daysbundled with RedEdge-P and 6Xsameincluded
DIY light sensorIf you built the budget rigAS7265x spectral triad at Mouser EUSparkFun€70
Reflectance panelThe reference every frame is scaled againstMAPIR calibration targetsame€150 to €400

The honest comparison: a DJI Mavic 3M is a calibrated five band system, in the air, for less than the price of a bare RedEdge-P. It is the right answer for a service business that wants to fly next month. Build the payload yourself when you need a band the integrated systems do not offer, or when you are putting it on a dock aircraft that has to fly unattended.

In-house software

StageWhat we runLicenceReplaces
Radiometric calibrationPanel plus DLS correction, per band, from the vendor's modelown code, micasense/imageprocessing as referencePix4Dfields, Agisoft radiometric workflow
Band alignmentFeature-based per-band registration, since the lenses are not coaxialOpenCV, BSD
MosaicODM multispectral mode, or per-frame georeferencing when a mosaic is not neededAGPLper-hectare processing
Indicesrasterio and numpy. It is arithmeticBSD
Baseline and changePer-field time series in COG plus a per-pixel z-score against the field's own historyown codethe entire "insights" tier of commercial platforms
RankingConnected components on the change layer, sorted by area times severityscikit-image, BSD
Scouting routeNearest-neighbour over the ranked patches, exported as GPXown code

The change detection is where the product is. Everything above it is commodity. Store every flight as a Cloud Optimized GeoTIFF keyed by field and date, and the "compare against the field's own history" step is a windowed read, not a pipeline.

Cost efficiency

  • Free satellite first, drone second. Sentinel-2 gives you 10 m NDVI every five days for nothing from the Copernicus Data Space. It cannot see a disease pocket, but it can tell you which of 40 fields changed this week, which is how you decide where to send the drone. This single decision cuts flying hours more than any hardware choice.
  • Do not buy a thermal camera for disease. It is tempting and it mostly measures water status. Spend the budget on calibration discipline instead.
  • One camera, many aircraft. The multispectral payload is the expensive part and it is not flying most of the time. Design the mount so it moves between airframes in under five minutes.
  • Weather data is free. The DWD publishes its open data, including model output, at opendata.dwd.de. Infection period models for septoria, blight and mildew run on temperature, humidity and leaf wetness, and they are what tell you to tighten the cadence. There is no reason to pay for that input.