This paper exists because of a citation, not an idea. Researching data sources for the all-of-Nairobi buildings pipeline, we came across prior academic work studying OpenStreetMap building-footprint completeness across global cities, which noted — in passing, as one data point among many — that Mathare, one of Nairobi's largest informal settlements, had no building footprints in OpenStreetMap at all at the time of that research. That's a striking claim, and having just pulled a full metro building dataset for an unrelated reason, we were in a position to check it against current, real data rather than take it on faith either direction.
In plain terms: every building on a modern open map got there one of two ways — a person traced it, or a computer spotted it in satellite imagery. This paper counts which was which, building by building, in two of Nairobi's informal settlements, and finds the difference between them has nothing to do with the settlements themselves.
Same city, same informality — opposite data histories. Green is a human who traced it; blue is a satellite that saw it.
What we actually queried
Overture Maps' buildings layer carries, for every feature, a sources field listing which upstream dataset(s) contributed it — OpenStreetMap, Google Open Buildings, or Microsoft ML Buildings, sometimes more than one per building where sources agree. We drew approximate bounding boxes around two informal settlements and counted buildings by source within each:
Kibera (36.780–36.795°E, -1.317–-1.308°N): 8,130 buildings. 6,706 (82.5%) from OpenStreetMap, 1,061 (13.1%) from Google Open Buildings, 363 (4.5%) from Microsoft.
Mathare (36.855–36.870°E, -1.265–-1.255°N): 5,061 buildings. Only 794 (15.7%) from OpenStreetMap, 3,546 (70.1%) from Google Open Buildings, 721 (14.2%) from Microsoft.
The claim about Mathare turns out to be dated but directionally correct — OpenStreetMap coverage there has grown from "essentially zero" to 16% since the research we found was published, but it remains the minority source by a wide margin, with ML-detected footprints supplying the overwhelming majority of what's actually on the map.
Why Kibera looks completely different
The gap between these two settlements is not explained by anything about the settlements themselves. Both are large, dense, informal, under-served by formal city infrastructure, and popularly cited in the same breath as examples of Nairobi's informal housing. What differs is their mapping history: Map Kibera, a volunteer OpenStreetMap project, has been actively training local residents to map their own settlement since 2009 — now a fifteen-year-plus, sustained, community-embedded effort specifically targeting that one settlement. Mathare has had no equivalent sustained campaign at comparable scale.
That is the actual variable. OpenStreetMap completeness, in an informal settlement, measures whether that specific place was fortunate enough to attract a dedicated volunteer mapping project — not its population, not its need for services, not how physically mappable its structures are from the ground or from imagery. Kibera is not "more mappable" than Mathare. It was mapped, deliberately, by name, for fifteen years, and Mathare mostly wasn't.
A blank spot on a volunteer map is not evidence of absence. It's evidence nobody was sent to look.
The part that should give anyone pause
If OpenStreetMap coverage tracks prior philanthropic and NGO attention rather than settlement characteristics, then any project that sources building data from OpenStreetMap alone is not measuring the city — it's measuring where past attention already landed, and reproducing that same unevenness downstream in whatever it builds next. A civic-tech tool, a disaster-response layer, a service-delivery planning map: built on OpenStreetMap-only data, each of these would systematically undercount exactly the places with the least prior institutional attention, which are disproportionately likely to be the places such a tool exists to help.
Data debt compounds exactly like financial debt: the places that start unmapped get built around, and every system built on the map raises the cost of being off it.
This is not a criticism of OpenStreetMap or of Map Kibera — both are real, valuable, and Map Kibera's fifteen years of sustained local mapping work is precisely the kind of ground-truth-accurate data no satellite model can fully replace (an ML footprint knows a rectangle exists; a Map Kibera contributor knows what it's called, whether it's a school, whether it floods). The point is narrower: an ML-detected footprint dataset does not wait for a volunteer campaign to arrive. It sees whatever the satellite saw, everywhere the satellite flew, which for Google Open Buildings and Microsoft's model is effectively everywhere. That's not a superior kind of data — it has its own failure modes, principally dense/informal roofing materials and heavy vegetation cover degrading detection confidence, and it carries none of Map Kibera's local knowledge (names, function, condition). But it does not have OpenStreetMap's specific bias toward wherever a human happened to already be paying attention.
A caveat, stated as plainly as the finding
We are not claiming ML-detected footprints are unbiased — satellite-imagery building detection has its own documented failure modes around dense informal roofing, cloud cover, and resolution limits, and a footprint with no name, no building type, and no local knowledge attached to it is a much thinner fact than a Map Kibera-traced one. The claim is comparative and specific: for a place with no dedicated mapping campaign, a merged dataset supplies most of what OpenStreetMap alone would simply not show, and that gap tracks attention, not reality.
What this changes, practically
For the Nairobi buildings pipeline this paper is a companion to, the implication was already acted on: the decision to use Overture's merged layer rather than OpenStreetMap alone was made before this specific comparison was run, on general principle. This paper is the evidence that the general principle was right, in a specific, checkable, real-numbers way, for the two Nairobi settlements it's cheapest to be wrong about.
Data & methods availability
Counts come from Overture Maps release 2026-06-17.0, whose per-building sources field lists the contributing dataset(s). Settlement extents are approximate hand-drawn bounding boxes (Kibera: 36.780–36.795°E, −1.317–−1.308°N; Mathare: 36.855–36.870°E, −1.265–−1.255°N) — adequate for source composition, which is a ratio, but not for authoritative building totals, which depend on where the boundary is drawn. Where a building lists multiple agreeing sources, the first-listed source is counted. Both caveats bias neither settlement relative to the other.
Companion papers
The pipeline that produced the underlying extract is documented in WP·from-cbd-to-all-of-nairobi; what we do about the attributes satellites can't see (height, use, condition) is in WP·estimating-what-we-dont-know.
