Tile-by-tile review vs. embedding search for archive-wide site hunting

Aerial view of dense forest canopy extending to the horizon beneath heavy clouds and distant mountains.

You've got one site. A suspected transshipment yard, a camouflaged pad, a compound with a distinctive wall layout. You already know it's a hit. What you don't know is whether five more like it are sitting somewhere in a collection area the size of a small country, unlooked-at because nobody's tasked a search for the other four yet.

That's the job this post is about: taking one known site and finding every other place in the archive that resembles it. Two ways to do that exist right now. One is scrolling. The other is querying an embedding space. They don't scale the same way.

What tile-by-tile review costs

Manual review means opening the archive, picking a grid, and paging through tiles at whatever zoom lets you actually see wall lines and roof shapes. For a single AOI that's a few hours of tedious but doable work. For a collection footprint that spans multiple provinces or a whole coastline, it's not a workflow anymore, it's a backlog.

The real cost is attention decay, not just hours. Tile two hundred gets the same scrutiny as tile twelve only if the analyst hasn't started skimming, and by the afternoon they're skimming. A look-alike site sitting at the edge of a tile, partly obscured by a tree line or shot at an oblique angle, is exactly the kind of thing a tired eye passes over. You don't find out what you missed until something else flags it months later, or until it shows up in someone else's report first.

There's also the archive problem. Continuous high-res collection means the pile of imagery to check keeps growing even while you're still working through last month's tasking. Tile-by-tile review was fine when the search area was one province and the archive was one season. It stops being fine once both of those scale up, and for most working search areas now, both already have.

Embedding-based similarity search skips the grid entirely. Instead of defining an area and paging through it, you mark the one site you already know is relevant, the reference tile, and the system compares its visual signature against every other tile in the search area. What comes back is a ranked list of places that share structure, layout, and texture with your query, not a fresh grid to re-scroll.

This is the core idea behind Broad Area Search: mark one place on the map, get back the other places across your search area that look like it, built on imagery embeddings run over high-res satellite tiles. It turns "search this whole region" from a scrolling task into a retrieval task.

The practical difference shows up in what the analyst's day looks like. Instead of covering ground, you're reviewing a shortlist. Instead of hoping you didn't blink past tile three-forty, you're deciding which of the twenty returned candidates deserve a closer look and which are false positives worth discarding. The archive doesn't get smaller, but the part of it you have to look at with your own eyes does.

It's also worth saying what this doesn't replace. A ranked candidate list still needs an analyst's judgment on each hit. Similarity search finds places that look like your reference site; confirming what they are, and whether that resemblance means anything operationally, is still the analyst's call. That judgment now gets spent on candidates instead of on blank tiles that never had anything to find.

When manual review still earns its keep

None of this argues for abandoning manual review outright. A small, well-bounded AOI where you already roughly know what you're looking for is still fast to eyeball directly, and there's no substitute for a trained eye catching something a query tile never would have surfaced, a structure that doesn't resemble anything you've flagged before. Similarity search is built for the opposite case: a search area too large to browse, anchored to one site you're confident is the right reference point.

That's the gap this capability is meant to close, and it's early enough in its life that we're looking for analysts willing to try it on a real search area and tell us where the returned matches hold up and where they don't.