Aircraft, drones or satellites for AI training data

The five variables that decide it

Every imagery source trades off the same five things: how much area it covers per capture, what ground resolution it reaches, when it can capture, how much control you have over the capture, and how the cost scales. Pick the source by the variable your product is most sensitive to.

Drones

A drone flies low and slow. That gives very fine ground resolution, down to centimetres per pixel, and cheap capture of a small site. The limits are structural: a battery lasts 20 to 40 minutes, the pilot must keep the aircraft in sight, and in most countries the airspace rules keep drones below a few hundred metres. In practice one drone flight covers hectares. A regional project means dozens of flights, several crews and weeks in the field, and the dataset arrives as a patchwork captured on different days under different light.

Drones win when the site is small, the detail must be extreme, or the capture is a one-off inspection.

Satellites

A satellite covers whole regions in a single pass and revisits on a fixed schedule. Archives go back years, which is valuable for change detection. The limits are the opposite of the drone's: resolution is bounded by the platform, the revisit date is not yours to choose, clouds can cost you the pass, and you cannot change the sensor or the angle. For a model that needs fresh imagery of a specific area on a specific week, at a resolution finer than the satellite offers, the archive does not help.

Satellites win when the area is huge, the cadence must be continuous, and moderate resolution is enough.

Manned aircraft

A manned aircraft sits between the two. It covers hundreds of square kilometres per flight, with hours of endurance, at a resolution you choose by setting the flying height. It flies on your date, over your area, with the sensor your project needs, including your own sensor if your product depends on it. The whole area is captured in one campaign, under consistent conditions, with overlap designed for photogrammetry.

The cost of an aircraft is operational, not capital. Someone has to plan the flight, get the permits, coordinate a certified operator and a pilot, run the sensors, monitor coverage during the flight, and process the data. That is the part technology companies without a local team do not want to build. It is also the part AeroStratus runs: we coordinate certified local operators in Chile and Spain, deploy a mobile operations base with its own connectivity, process on site and deliver into the client's systems.

Aircraft win when the area is regional, the resolution must be high, the date matters, and the data must arrive as one consistent dataset ready for a pipeline.

A short decision rule

• Hectares and extreme detail: drone.
• Whole countries at moderate resolution on a fixed cadence: satellite.
• Hundreds of square kilometres, high resolution, on your date, into your pipeline: manned aircraft.

Many products end up combining sources. Satellite archives for the broad baseline, aircraft campaigns for the high-resolution training and validation sets, drones for a handful of ground-truth sites. The mistake is choosing one source for reasons of habit and then discovering that the model is capped by it.