Terrapattern, the first open-access, neural net-powered reverse image search tool for satellite imagery, launches in alpha
“Why don't you click on the tennis court?” Golan Levin, an associate professor of art at Carnegie Mellon University, suggested. I was looking at a satellite image … Thanks: @ravichandrans25
Context & Ripple Effects
Terrapattern arrives three months after Google's PlaNet showed a neural network could outperform humans at guessing where a photo was taken — Levin's team at Carnegie Mellon flips the direction of that trick, letting anyone query satellite imagery by clicking on an example feature rather than typing coordinates. As the first open-access tool of its kind, it moves visual similarity search over geospatial data out of corporate labs and into the browser.
The alpha is also an early proof point for a research line that keeps extending: within four years, [[a:960463|convolutional networks are being used to comb satellite imagery for ancient burial sites across Russia, Mongolia, and China]], suggesting the search-by-visual-pattern approach generalizes well beyond Terrapattern's demo use cases.
First-order effects
- Researchers, journalists, and civic groups gain free access to search-by-example over satellite imagery — no ML expertise or paid API required — with Golan Levin's CMU team as the named provider.
Second-order effects
- Domain specialists adopt the same convolutional-search recipe for their own targets, as archaeologists later do for burial mounds, while startups apply adjacent visual-AI indexing to commercial verticals like OpenSpace's camera-captured construction-site records.
Third-order effects
- If open tools make satellite imagery trivially searchable, the value shifts from access to trust — a tension that surfaces a decade later when Google adds guardrails to Earth's image generation over deepfake satellite imagery concerns, making provenance of what these searches index the structural question.
The trend: Visual AI is turning geospatial imagery into something searchable by example rather than location, with the eventual bottleneck shifting from access to verifying what the imagery shows.