Aerial annotation

Aerial and Drone Image Annotation for Satellite and UAV Imagery

Overhead imagery breaks the habits of street-level labelling: objects are tiny, rotated and spread across huge images. Our rules are written for that view from the start.

Start a Pilot Explore Services

Compatible workflows

Tools we can work with

Roboflow
SuperAnnotate
Darkhorse
Roboflow
SuperAnnotate
Darkhorse

TL;DR

Large orthomosaics are tiled with overlap, and labels are mapped back to full-image and georeferenced coordinates without duplicates at tile edges.
Oriented bounding boxes follow the object's real heading, so parked cars, ships and rooftops are not padded with empty ground.
Agriculture labels separate crop, weed, bare soil and damage, with rules for overlapping plants and growth stages.
Thermal inspection labels mark hotspots and faults on panels, lines and roofs, kept apart from reflections and sun loading.

Imagery

Satellite imagery, UAV footage and orthomosaics

Aerial data arrives in very different forms: single satellite scenes, drone photo sets, stitched orthomosaics and drone video. The guideline starts by naming the source, because resolution and viewing angle change what an annotator can honestly see.

Resolution and visibility rules

Every class gets a minimum visible size. Below it, the object is either skipped or labelled as uncertain, depending on what your model needs. Annotators work at a fixed zoom range so that one person does not label objects another cannot see. Shadows, which often look more like the object than the object itself from above, are excluded from geometry unless a class says otherwise.

Off-nadir and stitching artefacts

Tall structures lean in off-nadir and oblique images, so the guideline states whether a building is labelled by its footprint, its roof or its full visible extent. In orthomosaics, stitching seams, ghosting and moving objects captured twice are flagged rather than labelled as two objects.

Tiling

Tiling and georeferenced coordinates for large images

Full orthomosaics are often too large to label or train on in one piece, so they are cut into tiles. Tiling is where duplicate and broken labels creep in if the rules are loose.

Overlap and edge rules

Tiles overlap so that objects cut by an edge appear whole in at least one tile. The guideline says which tile owns an object that crosses an edge, and how truncated parts are flagged. When tiles are merged back, overlapping labels of the same object are matched and joined, and checks flag any object that appears twice.

Pixel and map coordinates

Labels are stored in tile pixel coordinates and can be converted to full-image pixels and, where a GeoTIFF transform and coordinate reference system are supplied, to map coordinates. We keep the transform alongside the labels so your team can rebuild the same geometry in GIS tools. Delivery can include GeoJSON for mapped features next to the training formats.

Geometry

Oriented bounding boxes, polygons and small objects

From above, most objects are rotated relative to the image grid. An axis-aligned box around a diagonal ship or a parked car is mostly ground.

Oriented boxes and heading

Oriented bounding boxes are defined by centre, width, length and angle, or by four ordered corners. The guideline fixes the angle convention and which side counts as the front, for example the bow of a ship or the front of a vehicle, so headings are consistent. Where the front cannot be seen, the box keeps its shape and the heading is marked unknown.

Buildings, fields, water bodies and roads are usually polygons. Small and dense objects, such as cars in a car park or solar panels in an array, get explicit rules for touching instances and for when a group is labelled as one region.

Box, polygon and mask rules for still images are covered on our image annotation page.

Agriculture

Agriculture: crops, weeds, growth stages and field damage

Field imagery is mostly one plant next to another plant. The class list and overlap rules decide whether the model can tell them apart.

Crop, weed and condition labels

Crop rows, individual plants, weeds and bare soil are labelled with masks or polygons, and the guideline states how to handle overlapping leaves and plants at the edge of a row. Condition attributes such as stress, disease symptoms, lodging, flooding or hail damage are defined with example images, and growth stage is labelled only when it is visible, not assumed from the date.

Multispectral and change

When bands beyond visible light are supplied, annotators can label on a chosen band combination named in the guideline. For change detection, the same area is labelled at two dates under one taxonomy, and each change is tagged with what changed, not only where.

Inspection

Thermal inspection of solar, power lines and roofs

Thermal drone surveys are used to find faults before they fail. The labels must separate real faults from heat that has an ordinary cause.

Hotspots and fault classes

On solar arrays, annotators label hotspots, cell and module faults, string outages and soiling as separate classes, with the module or panel identified where the layout allows. On power lines and roofs, labels mark hot joints, damaged insulators, missing tiles and moisture patterns. Reflections of the sky or sun, and heat from normal operation, have their own not-a-fault rules.

QA follows our two-stage process: annotators self-check against the guideline, then an independent reviewer audits each batch for missed small objects, wrong headings, tile duplicates and fault classes that drift over time.

For drone video tracking and inspection events, see our video annotation page.

Aerial and drone annotation FAQ

How do you label objects that cross tile edges?

Tiles overlap, and the guideline says which tile owns an object on an edge. When tiles are merged, matching labels are joined and checks flag any object that appears twice or is cut in half.

Can you deliver labels in map coordinates?

Yes, when you supply a georeferenced image with its transform and coordinate system. Labels are kept in pixel coordinates for training and can also be delivered in map coordinates, for example as GeoJSON.

When should we use oriented bounding boxes?

Use them when objects are elongated and rotated, such as vehicles, ships, aircraft and rooftops seen from above. The guideline fixes the angle convention and front side so headings stay consistent.

How do you avoid false faults in thermal imagery?

The guideline separates real faults such as hotspots and failed modules from reflections, sun loading and normal operating heat, with example images for each. Unclear cases are flagged for your inspection team.

Send us a tile from your largest map.

Start with a Free Calibration Pilot