Drone images of real cities, licensed for model training

Aerial imagery for the Physical AI teams building world models and spatial intelligence.

Millions of images across hundreds of U.S. and Canadian cities, captured at 2.8 cm or sharper and licensed to train models and deploy them commercially.

See what's captured where your model needs to work

List the cities your model will operate in. We'll reply with what's already captured in each one.

Millions of images, already captured across hundreds of U.S. & Canadian cities

  • Miami, FL
  • Vancouver, BC
  • Chicago, IL
  • Halifax, NS
  • Phoenix, AZ
  • Toronto, ON
  • Seattle, WA
  • Ottawa, ON
  • League City, TX
  • Calgary, AB
  • New York, NY
  • Montréal, QC

Good aerial data for physical AI training is hard to get

Unclear rights, costly collection and messy imagery slow teams down before training even starts.

  • Data provenance risk

    Unclear image rights can stall a launch, or force you to retrain from scratch.

  • Data collection cost

    Collecting the data yourself burns runway before training even starts.

  • Curation and coverage gaps

    Engineers clean imagery instead of training, and top-down satellite views miss the angles your model needs.

Spexi archive data is the solution

One licensed, training-ready archive of accurate, measurable drone imagery, with an aerial view of every city block in its coverage.

A 12-month calendar showing large captures in April and October, bracketed as the 6-month maximum, plus smaller update captures in January, June, August and December.
01

The aerial view your model is missing.

Multiple angles and altitudes at 2.8 cm or sharper, to pair with the ground-level data you already have.

Straight-down view of a building's tile roof with skylights and rooftop solar panels
Orthomosaic
Angled view of the front of a brick building with a tile roof and rooftop solar panels
Angled imagery
Wide aerial view of a brick building with a tile roof, parking lots and a tree-lined street
360° views
02

Every capture type in one archive.

Top-down and angled images, 360° panoramas, plus orthomosaic-ready and 3D-reconstruction-ready subsets.

Five drone images of the same city block stacked over time, from 24 months ago to today, showing a vacant lot becoming a finished subdivision.
03

Query the training mix. No sorting required.

Every image is captured to the same standard with the same metadata, so you build the mix with queries.

Three ways to access imagery: Spexi Worldviewer for non-GIS teams, an ArcGIS layer for GIS teams, and the Image API for developers and IT.
04

One contract clears every image.

One licence with rights to train models and deploy them commercially, so legal reviews a single contract.

How it works

Every image is a structured record

The Spexi pilot network flies standardized, automated flight plans. Each image enters the archive with its footprint, camera pose and timestamps.

  1. Search by place and angle

    Filter by bounding box, polygon, capture type, date, heading, pitch and altitude.

  2. Metadata on every image

    Footprint, heading, pitch, roll, altitude, camera position and timestamps.

  3. 2.8 cm resolution

    Ground sample distance per pixel, or sharper.

A drone shines a green capture beam onto a 3D model of a waterfront city, with glowing hexagons outlining Spexigons on the ground.

Resolution

See the detail your model trains on

An airplane flyover capture of a building drop-off lane: the lane markings, curb and plantings blur together.

Airplane flyover

Typical 5–7 cm per pixel

Spexi drone capture of the same drop-off lane: lane arrows, the painted DROP OFF ONLY text, a speed bump, rooftop equipment and individual shrubs are distinct.

Spexi drone capture

2.8 cm per pixel

  • Angled view of a subdivision with new homes under construction outlined

    3D reconstruction

    Multi-view captures with 70% overlap support generalizable reconstruction and novel-view synthesis.

  • Straight-down view of a backyard pool and rooftop solar panels outlined

    City-scale geolocalization

    Match ground and aerial views across hundreds of cities, with camera pose on every image.

  • Angled view of a residential road and sidewalk outlined

    Aerial foundation models

    Pre-train on millions of standardized images from every season and angle.

  • Angled view of a sports field and running track outlined

    Long-tail object detection

    Query rare scenes like backyard pools, construction sites and complex intersections.

  • Straight-down view of a utility pole with lines running into tree canopy outlined

    Cross-city generalization

    Test how models hold up across climates, regions and urban layouts.

  • The same street in February and June 2025, after a wildfire and after cleanup

    Temporal and seasonal robustness

    Captures from 2024 to today show the same places across seasons.

Frequently asked questions

Still have a question? Tell us the cities your model needs and we'll show you what's already captured.

What does the licence let us do?

Train models on the archive and deploy them commercially for the term of your licence. Raw images may not be resold.

Is the licence exclusive?

No, it's non-exclusive. Its scope is set by the archive size at purchase, and it can be amended as new imagery is added.

What metadata comes with each image?

Footprint polygon, capture type, heading, pitch, roll, altitude above ground, camera position, capture and processing timestamps, and tags.

How is the data delivered?

By bulk transfer as JPG files, with a metadata file you can filter on any field.

What about privacy and data residency?

Faces and licence plates aren't identifiable at the archive's resolution. Flights follow FAA and Transport Canada rules, and U.S. customer data stays in the U.S., Canadian data in Canada.

What could your model learn from the full archive?