Mark one site. Get back every place across your search area that looks like it.

Similarity search over high-resolution satellite imagery, in early access. You give us the query tile. You get a ranked candidate list to triage.

Join the pilot listSee how retrieval works
GSD
High / 0.5–2 m
Cadence
Continuous
Platform
High-res satellite
Spectral
RGB
Schematic search grid with one query tile and two look-alike sites
Fig. 1Schematic, not imagery. Black: the query tile. Red-orange: two look-alike sites in the search area.

One known site, and an archive too big to browse

You have a site you are sure about. It came out of a report, a tasking, or a tip somebody confirmed on a good pass. Now the question from upstream is where else there are others like it, across a region or a whole country.

Today that is a manual broad area search. You set the search area in your imagery viewer, cut it into a grid, pin the known site in a side window and start down the rows. After enough hours your attention drifts. You reach the end of a grid square and cannot say honestly whether you saw every candidate or only the ones that stood out. A colleague working the next squares over carries a different idea of what counts as close enough.

A dry plain running flat to a far mountain range, one dirt track crossing it
Open ground, photographed from the road. Scenery, not imagery from the service.

The archive keeps growing, and the area you can cover by eye does not. The search ends when the suspense date arrives, and the area is rarely done by then. A product that says "none found" then means none in the tiles you reached.

This offer changes the first pass. Retrieval ranks the whole search area against your query tile, and you start reading from the top of the list instead of the top-left corner of the grid. The judgment about what each candidate is stays with you.

What early access gets you

This is a pilot. Nothing here is a shipped product, and each item below is what the pilot is set up to produce.

2.1Query tile in

Query-by-example search

You mark one site on the map, or send a location, and that is the whole query. There is no training set to label and no model to configure before the first search.

2.2Ranked list

Ranked candidate list

Each candidate comes back with its location and its similarity rank. A rank tells you which locations look closest to your query tile. A confidence score would claim more, so the list carries none. It is meant to open in the viewer you already use.

2.3Search area

Search-area scoping

You define the region to search over. We tell you what imagery exists across it before anything is ranked, so you know the coverage the ranking sits on.

2.4Pilot input

Pilot feedback loop

Pilot users tell us which candidates were worth a closer look and which were noise, and that input shapes what the retrieval surfaces next.

How it works

Step 1: Send the query tile

Mark the known site on the map, or send its location, and outline the search area you care about.

Step 2: Imagery is indexed

High-resolution satellite imagery over the search area is turned into embeddings, numeric summaries of what each patch looks like.

Step 3: Similarity ranking

The embedding of your query tile is compared with the rest of the search area, and every location is ranked by how close it comes.

Step 4: Candidate list back

You receive the ranked list for review. Read from the top, discard the false positives, and tell us which ones were worth the look.

Where it fits in your workflow

It sits before your own exploitation, ahead of the step where you open a candidate and read it. Use the list to decide where to spend your time first, then do the analysis the way you always have. The output is a lead list for your triage, and the call on each location is yours to make and sign.

Limits

These are the ways a similarity ranking can mislead you. The page states them up front because the list is only useful if you read it with them in mind.

  1. L1

    A ranking, not an answer

    Retrieval sorts by visual similarity. It does not find every match, and it does not promise that the top of the list is right. Expect false positives, and expect real look-alikes further down than you would like.

  2. L2

    0.5–2 m is the floor

    At this resolution a site is matched on its overall layout and footprint: how buildings, pads, roads and surroundings sit together. Fine sub-meter detail is not what it matches on, and small equipment or single vehicles will not decide the ranking.

  3. L3

    Only where imagery exists

    Coverage depends on what has been collected over your search area. Where there is no usable image, or cloud sits over the site, nothing can rank there, and the list cannot tell you about it.

  4. L4

    RGB only

    The imagery is visible-band. Anything that only shows in other bands is out of reach.

  5. L5

    Looks alike is not the same as is alike

    Two sites can share a footprint and serve different purposes. The list gives you resemblance in the picture. It has no way to tell you function.

  6. L6

    Early access

    This is a pilot, so the interface, the outputs and the way results are ranked will change as pilot users push on them.

Who it's for

  • Imagery and GEOINT analysts with one known site and a wide search area, doing the search by eye today.
  • Teams that want a ranked lead list to triage before spending analyst time on a full manual pass.
  • Groups willing to run a pilot and say plainly which results helped.

Who it's not for

  • Teams that require a complete inventory with a guarantee that nothing was missed.
  • Teams looking for an automated target list. This produces candidates for an analyst to review.
  • Work that depends on fine sub-meter detail, such as identifying individual vehicles or equipment.
  • Buyers who want a finished, supported product today.

Questions from analysts and their leads

Is this a finished product?

No. It is an early-access pilot, and joining the list is how you get a say in what it becomes.

Does it find every match?

No. It ranks the search area by similarity and gives you a candidate list. It is not exhaustive and nothing about it is guaranteed.

What resolution does it search at?

High-resolution satellite imagery, 0.5–2 m. It matches on layout and footprint, not on fine sub-meter detail.

Do I need to label training data?

No. One query tile is enough to start.

How do we check it against a site we already know?

Pick a query tile whose look-alikes you already know about somewhere in the search area and see where they land in the ranking. That is the plainest test, and we will run it with you in the pilot.

Will it fit the tools we already use?

The list is meant to open in the imagery viewer or GIS your analysts already work in. Which formats and systems we support is part of what the pilot settles.

Who owns the results, and what happens to our query?

Ownership, handling and confidentiality terms for your query and the results are written into the pilot agreement before anything runs. Say so when you join and we will start there.

How is it priced?

Pilot scope and terms are set per group, and a specialist will confirm them with you. There is no price on this page.

Tell us your query tile and search area.

We will talk through how a pilot would run.

Join the pilot list