Every building on Earth, on one map — with its measurements, its height, and its neighbours.
One HTML file. No server, no database, no build step.
Imbaba, Giza. Imagery © Esri World Imagery. Building outlines © Google, Microsoft and Overture contributors.
- Open it — it lands over Imbaba, Cairo.
- Click any building. Its record opens: area, perimeter, compactness, how many buildings physically touch it, whether it fronts a street.
- Switch Buildings to "In selection", drag a rectangle. You get the neighbourhood's density, coverage and block complexity.
- Press "Download all". A ZIP lands with the two-colour mask, the instance mask, the footprints as GeoJSON, the statistics as CSV, and world files so it opens in QGIS in the right place.
Nothing is uploaded and nothing is stored. Everything above is computed in your browser.
Click any building and read its record: footprint area, perimeter, compactness, the exact number of buildings physically touching it, height in 2023 and 2020, an estimate of storeys, and its share of the built volume around it.
Travel in time. The imagery slider moves through Esri's archive from 2014 to 2026. Where two years turn out to be the same photograph, the app says so rather than letting you wonder why nothing changed.
Find the best picture, not the newest one. The archive holds 196 releases and the newest is a publication order, not a quality order. "Sharpest here" searches it for this ground — over Kibera it finds imagery 1.6× sharper and twice as well placed as the one on screen.
Go anywhere. Search any place on Earth, or use the find-me button when you are standing in the area you are surveying. The map position lives in the URL, so a view can be sent to a colleague as a link.
Read it in English. Street and place names appear in English wherever the data has one, and are romanised character by character where it does not, so Greek, Cyrillic, Arabic, Persian, Hebrew, Thai and Devanagari all render in Latin script.
Measure a neighbourhood. Drag a rectangle and get its count, density, footprint distribution, how much ground is actually under a building, and block complexity — the published metric for how many parcels you must cross to reach a street, from Infrastructure deficits and informal settlements in sub-Saharan Africa (Nature, 2025). k = 1 is a planned block; k = 5+ is a severe access deficit.
Take it away as a deliverable. One ZIP: a two-colour mask with touching buildings
separated by a one-pixel seam, a colour instance mask on the same grid, the footprints
as GeoJSON, every figure as CSV, and world files plus .prj so it lands correctly in
QGIS. Rendered in EPSG:3857, to match the imagery it was traced from.
Bring your own imagery. A GeoTIFF lands at its own coordinates — WGS84, Web Mercator or any UTM zone, read from the file's own header and reprojected. A drone JPEG is placed from its EXIF GPS. A plain photograph you place, then snap to the imagery by correlation. The image workspace describes the image and nothing else — the map's outlines switch off while it is open and are handed back untouched when you remove it.
Bring your own polygons. GeoJSON, KML and CSV (points or WKT) are scored building by building against the published data — which agree, which you have and they do not, which they have and you do not. Matched by centroid-inside-polygon, not by overlapping boxes.
Read the picture itself. A public 15 MB segmentation model, run in the browser on your own image — the one place this app looks at pixels rather than re-serving what someone else traced. Its result is always scored against the published footprints and never shown alone, because its quality on dense informal fabric is unmeasured.
Works on a phone, a tablet and a desktop, in a light or a dark theme.
| Measured | footprint geometry, area, perimeter, compactness, touching neighbours, counts, density, coverage, block complexity, the source's own confidence |
| Estimated | height — from a ~76 m grid, so it describes the block rather than the individual roof, and reads about 2 m low against finer data; floors — height ÷ 3.0 m, an estimate built on an estimate; population share — each building's share of the local built volume, to multiply by a total you trust, and never a headcount |
Every estimated figure carries that caveat in the panel itself, next to the number.
Two open datasets often draw the same roof. At zoom 17, the share of Overture polygons sitting on a building the other source had already drawn was 74% over Imbaba, 90% over Cairo and 71% over Lagos — which is why buildings appeared to carry several outlines stacked on them. Neither source can simply be dropped, because which one is better flips by place: 634 buildings against 208 over Imbaba, but 32 against 253 over Paris, and none at all over Beijing. The explorer draws the denser product for the current view, so no roof is outlined twice, and the other is one click away under Sources — where turning both on is an explicit choice to see where they disagree.
Position is bounded by the imagery, not by the outline. Measured across 32 places: aerial-survey cities (Johannesburg, New York, Berlin) are accurate to 0.4–0.6 m, and satellite coverage — most of the world — to 8.47 m. A footprint traced from a picture inherits that picture's accuracy however precise its outline looks. The panel says so, per area, next to the numbers it bounds.
Some imagery years are the same photograph. Comparing tile bytes over Imbaba, 2014, 2015 and 2016 are identical, and so are 2025 and 2026 — thirteen year options, ten real captures. Which years collide depends on where you are, so it is resolved per tile.
Building outlines date from about 2023 while the imagery may be older or newer. On a 2014 basemap you are looking at today's buildings over yesterday's ground, and the panel warns you when the two drift apart.
| Layer | Source | Licence |
|---|---|---|
| Buildings | Google Open Buildings + Microsoft, combined by VIDA | CC BY-4.0 |
| Buildings | Overture Maps | ODbL / CC BY-4.0 |
| Height, density | Microsoft Building Density | CDLA-Permissive-2.0 |
| Imagery | Esri World Imagery Wayback | Esri terms |
| Roads, divisions | Overture Maps / OpenStreetMap | ODbL |
| Place search | Nominatim / OpenStreetMap | ODbL |
| Segmentation model | geobase, via geoai.js | MIT |
The About the data panel in the app carries the same credits, and so does the README inside every export. Keep them in any copy you publish or redistribute. Full detail, with links and the software licences, is in ATTRIBUTION.md.
Requires an internet connection. Nothing is bundled — every layer streams from its publisher at the moment you look at it. That is what keeps one file covering the whole planet, and it also means the app cannot work offline.
Two libraries are fetched only when they are needed rather than on every page load: geotiff.js when you click a building or open a GeoTIFF, and ONNX Runtime Web plus the 15 MB model only if you ask it to read your own picture.
Built at NARSS (National Authority for Remote Sensing and Space Sciences, Egypt).
The explorer is the delivered product of a much larger study on segmenting individual buildings in dense informal settlements — the case that breaks most methods, because neighbouring buildings share walls and there is no gap between them to detect. That work involved fine-tuning segmentation models on Kaggle, testing several instance-separation approaches against one another, measuring the resolution ceiling of freely available imagery, and validating against independent building footprints.
Two results from it shaped this app directly, and are worth stating plainly:
- A fine-tuned model added about 9.5% over the public building data, at roughly 50% precision — and separated touching buildings worse than the public data already does. At 0.26 m per pixel, the bottleneck is not the model.
- 0.26 m per pixel is the ceiling for free imagery over the study area. Esri matches Google, Bing and Yandex are worse, there is no genuine zoom 20, and multi-frame super-resolution did not recover detail.
Which is why this explorer is built on the best available public data rather than on a model — a conclusion reached by measurement, not assumption.
The research repository — experiments, the full findings log, training notebooks and validation — is maintained separately and is not public.
This repository is the working copy. index.html originated from a generator in the
private project repository, but it is edited directly here now and the history is commits
against this file.
git add index.html && git commit -m "..." && git push origin mainGitHub Pages rebuilds within a minute or two.
⚠️ Do notgit subtree pushover this repository from the private side. Work now exists only here, and a subtree push would silently overwrite it. If the generator is ever re-run, merge this file into it rather than over it.
There is no build step and no test runner in this repository — it is one document. Open
index.html from any static server and drive it:
python3 -m http.server 8765 # then open http://localhost:8765/index.htmlIt is one static file. Any web server works — drop index.html anywhere and open it.
There is nothing to configure and nothing to keep running.
The viewer code is MIT — see LICENSE. The map data it displays is not covered by that grant; each layer keeps its own licence, listed in ATTRIBUTION.md.
A CITATION.cff is included, so GitHub's "Cite this repository" button produces a formatted reference. If you use the explorer or its outputs in published work, please cite it.