Build guide

On this page 6

How to build each part of this platform without renting it by the hectare.

Everything here follows one rule: the hardware is a commodity and the software is not. Anyone can buy a drone. The thing that normally turns a drone business into a thin reseller is the per-hectare licence on the software that turns imagery into a decision, and none of that software has to be rented.

The three cost lines, in order of size

LineTypical commercial routeIn-house route
Processing and analyticsPer hectare per year, forever, rising with your best customersOne GPU box, electricity
Aircraft and sensors€5,000 to €40,000 per rigThe same, or less if you build airframes
Compliance and operationsRegistration, insurance, training, maintenanceIdentical either way. This one you cannot engineer away

The first line is the one that decides whether the business works at scale, and it is the one this guide is mostly about.

Pages

PageCovers
AirframesBuy or build, the four aircraft classes, parts lists
SensorsRGB, multispectral, thermal, LiDAR: what to buy at each budget
PositioningRTK, correction services, and why a base station is usually the wrong purchase
ComputeOn-board, ground station, and the processing box
DocksUnattended operation, buy versus build
PayloadsHoppers, spray systems, deterrents, gimbals
SoftwareThe whole in-house stack, in five pages
CostsFour complete build tiers with running costs
SuppliersWhere to buy, EU and US
RegulationEU and German rules that change what you can build

The build order that works

Do not start with the aircraft. Start with the thing that produces a deliverable.

  1. Process somebody else's imagery. Take a public dataset or a single flight from a rented drone and run it end to end: stitch, detect, cluster, export a prescription. If you cannot produce a file a sprayer terminal accepts, no aircraft will help.
  2. Buy one boring aircraft. An integrated commercial multirotor with a calibrated multispectral payload. Fly paid work with it.
  3. Build the second aircraft for the payload the market does not sell you cheaply, usually thermal for wildlife searches or a high resolution RGB rig for weed control.
  4. Add a dock once one customer's cadence justifies it.

Every step in that order produces revenue before the next one costs money.

What "in-house" does and does not mean

It means: no per-hectare software licence, no vendor cloud holding your customers' field records, no analytics tier you cannot audit, and no model you cannot retrain on your own fields.

It does not mean: writing a flight controller, a photogrammetry engine or a deep learning framework. ArduPilot, PX4, OpenDroneMap, COLMAP, GDAL, PyTorch and ONNX Runtime are mature, free, and better than anything a small team will write. Using them is the in-house route. The custom work is the layer above: the calibration discipline, the models trained on your own fields, the clustering into a prescription your customers' machines can follow, and the record that ties every figure back to the flight that produced it.

Licences to decide before you write integration code

This is the one piece of legal homework that will actually bite.

ComponentLicenceConsequence
ArduPilotGPL-3.0Modifications distributed with the aircraft must be published. Using it unmodified is fine
PX4BSD-3Permissive. Often the right choice for a commercial product
OpenDroneMap, WebODMAGPL-3.0Network use of a modified version triggers source obligations. Run it unmodified as a batch job, or use the permissive stack
COLMAPBSDPermissive
OpenMVGMPL-2.0File-level copyleft, workable
Ultralytics YOLOAGPL-3.0 or commercialThe most common expensive mistake in this field
RF-DETR, MMDetection, torchvisionApache-2.0 / BSDPermissive. Start here
GDAL, PROJ, rasterio, OpenCVMIT / BSDPermissive
MinIOAGPL-3.0Consider SeaweedFS (Apache-2.0) or plain S3-compatible hosting

None of these are reasons to avoid open source. They are reasons to pick the permissive option before the integration exists, because swapping a detector after eighteen months of labelled data and tuning is not a weekend.

A note on the prices in these pages

Every hardware table gives an indicative figure and both an EU and a US source. The figures are 2026 estimates and they move: sensor pricing in particular has fallen steadily and airframe pricing has not. Check the vendor's current price before you build a quote on anything here, and treat the tables as a shopping list rather than a budget.