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.1 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.
Same city, same informality — opposite data histories. Green is a human who traced it; blue is a satellite that saw it.
1. What a two-settlement comparison can carry, and what it can't
This is a narrow, two-settlement empirical check, and its evidence reaches only as far as stated here.
What it argues from. A direct query of Overture Maps' per-building sources field for two hand-drawn bounding boxes (Kibera, Mathare), decomposing each settlement's building count by contributing dataset. This is reproducible against the stated Overture release and is the paper's strongest evidence.
What it argues from, for the historical claim. A 2023 peer-reviewed study's reported observation about Mathare's historical OpenStreetMap coverage, checked against current data rather than simply repeated.1 A 2017 account, read directly, of Map Kibera's founding and methodology.2 An unreviewed 2025 preprint comparing building-dataset geometric quality across five cities including Nairobi, read directly and flagged explicitly as not yet peer-reviewed.3
What it does not argue from. It does not argue that two settlements are a representative sample of Nairobi's dozens of informal settlements, or of informal settlements generally across Africa. It does not argue that OpenStreetMap or ML-detected footprints are more accurate in any positional sense — only that they differ sharply in source composition, which is a narrower and more defensible claim. It does not claim the mechanism identified here (volunteer-mapping-campaign presence) is the only driver of coverage variation; it is the one this paper found evidence for in exactly two cases.
What would refute it. Given in full in §9. The sharpest version: if a third Nairobi settlement with no comparable volunteer-mapping history is found to have OpenStreetMap coverage similar to Kibera's 82%, the mechanism this paper proposes — that a dedicated campaign, not settlement characteristics, explains the gap — would need withdrawing rather than qualifying.
2. 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.
3. The source behind the Mathare claim, checked rather than repeated
The claim we started from comes from a 2023 peer-reviewed study of global OpenStreetMap building-data completeness — the same study already used elsewhere in this series for its funding-gap and humanitarian-mapping findings.1 That paper reports Sub-Saharan Africa's OpenStreetMap building completeness at 30%, "slightly higher than the global mean" of 24%, and — the figure directly relevant here — finds that more than half of all building edits in Sub-Saharan Africa, 51.1%, came from organised humanitarian mapping campaigns rather than ongoing institutional or individual contribution.1 That is the aggregate version of exactly what §2's two-settlement comparison finds concretely: coverage in this region tracks organised campaign attention, not population or need, and Kibera versus Mathare is one checkable instance of that aggregate pattern rather than an exception to it.
4. 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, and it is worth being precise about that history rather than citing it as received wisdom.
Map Kibera's own account of its founding states plainly: Kibera "was a blank spot on the map until November 2009, when young Kiberans created the first free and open digital map of their own community."4 The mapping itself — walking the settlement's paths and alleys with consumer GPS units after two days of training — was done by Kibera residents, not by outside researchers; the facilitating organisation, GroundTruth, co-founded by Erica Hagen and Mikel Maron, provided training, equipment and an initial platform.2 An eight-year retrospective on the project, commissioned by the Making All Voices Count research programme and written by Hagen herself, documents this as a sustained effort — not a single mapping event but a continuing project spanning digital mapping, citizen media and open-data work across education, security and water, concluding that "open, community-based data collection can lead to greater trust, which is sorely lacking in marginalised places."2 Mathare has had no equivalent sustained campaign at comparable scale or duration.
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 its own residents, trained and equipped for the purpose, for fifteen years running, 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.
It is also worth being specific about what the original 2009 effort actually produced, because "mapped" can mean several different things and this paper's claim rests on the richer one. Within roughly a month, the initial team of trained residents had traced not just building footprints and road alignments but the points of interest that make a map locally useful — water points, toilets, schools, pharmacies — categories the mappers themselves chose, not categories imposed by an outside research agenda. That is a materially different kind of map than a building footprint detected by a satellite model, which knows a rectangle exists and nothing else about it. The gap this paper measures in §2 is only the footprint-count version of a much larger gap in what is actually known about each settlement, and Kibera's fifteen-year head start widens that larger gap even where, as in Mathare, ML-detected footprints have since closed the narrower one.
5. 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, resident-led 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.
This is not a hypothetical concern specific to building footprints. The same logic governs any civic dataset built from volunteer contribution rather than systematic survey: a pothole-reporting app's coverage tracks who owns a smartphone and bothers to report; a crowdsourced water-quality map tracks where an NGO ran a training session; a disease-surveillance layer built from citizen reports tracks literacy and trust in the reporting institution as much as disease incidence. In each case, the gap in the data is not randomly distributed across the city — it falls disproportionately on the places with the least prior institutional or philanthropic attention, which are frequently the same places a given civic tool exists to serve. §2's Kibera/Mathare comparison is one checkable instance of a pattern this paper believes generalises well beyond building footprints specifically, though — consistent with §1's scope statement — this paper has not tested that broader generalisation directly.
6. A second, independent check on source reliability
Section 2's comparison is about coverage — how much of each settlement each source contributes. A separate, more recent question is whether the ML-detected sources that supply most of Mathare's buildings are themselves geometrically reliable in a city like Nairobi, independent of the settlement-coverage question entirely.
A 2025 preprint by Okyere, Lu and Brunn, comparing Google Open Buildings and Microsoft Building Footprints against reference data across five cities — Accra, Nairobi, Caracas, Berlin and Houston — reports, for Nairobi specifically, a moderate intersection-over-union agreement for Microsoft's data (an average IoU of 0.52, with 41.3% of Microsoft polygons overlapping a reference polygon) and a lower figure for Google's data (an average IoU of 0.40, 31.6% overlap), alongside the highest angular geometric deviation (1.42°) the paper measured for Microsoft's footprints across any of the five cities.3 The paper frames this in the context of the "global socio-economic divide" it finds reflected in building-size distributions, and reports that structured cities (Houston, Berlin) show higher alignment and completeness than cities "characterised by informality" such as Accra and Caracas — though the paper does not break its Nairobi results down by named settlement, and does not specifically examine Kibera or Mathare.3
This source's limits are worth stating honestly rather than letting it quietly borrow more authority than it has earned. It is an arXiv preprint, not yet peer-reviewed, and I found no evidence it has been published in a journal as of this writing. Its relevance to this paper is narrower than §2's own settlement-level finding: it establishes, independently, that the ML-detected sources this paper credits with filling Mathare's coverage gap are not themselves perfectly reliable in Nairobi specifically — a real caveat to §5's framing, taken up directly in §9.
The paper's own framing is also worth stating directly rather than only extracting its numbers: it interprets the pattern across all five cities — stronger alignment in Houston and Berlin, weaker in Accra, Caracas and (to a lesser degree, on its figures) Nairobi — as a reflection of a broader "global socio-economic divide" encoded in building-size distributions and urban morphology, not a Nairobi-specific quirk of either vendor's model. That framing is consistent with this paper's own argument, but it is a different claim from §2's: Okyere et al. are describing a geometric-reliability gradient that tracks a city's general level of formal urban structure, while this paper's §2–§4 describe a coverage gradient that tracks a specific settlement's mapping history. The two gradients may well be related — informal construction is exactly the "uneven urban fabric" the preprint says both ML models struggle with — but this paper does not have evidence that they are the same gradient, and treats them as two separate, mutually reinforcing cautions rather than collapsing them into one.
7. 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 §6's independent check confirms Nairobi is not a city where either major ML source achieves strong geometric agreement with reference data. 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.
8. 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.
9. Objections I take seriously
"Two settlements is not a sample; it's an anecdote with numbers attached." This is the most important limitation of this paper and §1 states it directly rather than wait for a reader to find it. Kibera and Mathare were chosen because one has the most famous volunteer-mapping history in the region and the other was the specific subject of the claim that started this paper — not because they were randomly sampled from Nairobi's dozens of informal settlements. The mechanism this paper proposes (campaign presence, not settlement characteristics, drives coverage) is well-supported for these two cases and consistent with the aggregate Sub-Saharan Africa figures in §3, but it has not been checked against a third, fourth, or fifth settlement in this paper.
"If the ML sources filling Mathare's gap are themselves only moderately reliable (§6), the paper may be overstating what the 'merge' actually buys." Fair, and I think §6 is the right place to have raised it rather than omit it. The claim this paper defends is about presence — does a building exist in the dataset at all — not geometric precision. Okyere et al.'s IoU figures speak to how well a detected polygon's shape and position match a reference, which is a different and, for many downstream uses, less urgent question than whether the building is represented at all. Both questions matter; this paper only answers the first one.
"Flux has a commercial interest in this conclusion." Flux sources building data from the same merged Overture layer this paper studies, for the Nairobi digital twin it sells. A finding that validates merging multiple sources over relying on OpenStreetMap alone is convenient for a company that has already made that architectural choice. The mitigation is that this paper's own §6 introduces an independent, less flattering finding — that the ML sources Flux's pipeline leans on for exactly the places OpenStreetMap misses are not themselves highly reliable in Nairobi — rather than omitting it because it complicates the headline story. A reader can check both the settlement-level source-composition numbers and the independent geometric-quality preprint without relying on Flux's framing of either.
"Crediting 'young Kiberans' rather than the outside NGO that trained them risks either over-crediting local agency or erasing the outside investment that made it possible." I tried to state both facts rather than pick one: Map Kibera's own public account credits residents as the mappers and names a specific month they started; Hagen's own retrospective, written by one of the facilitating NGO's co-founders, documents the training, equipment and platform GroundTruth provided. §4 cites both rather than flattening the history into either "local people did everything" or "an NGO did everything."
"This paper measures source composition, which is a count of polygons by origin — it does not measure whether the resulting map is actually used, trusted, or useful to anyone in either settlement." This is correct and worth stating as a limitation rather than disputed. Hagen's retrospective speaks to use and trust for Kibera specifically, built over eight years of the project's own fieldwork and interviews; this paper has no equivalent evidence for Mathare, and does not claim that Mathare's ML-majority map is used, trusted, or even known about by Mathare residents at all. A dataset existing in Overture and a dataset mattering to the people it describes are different claims, and this paper only makes the first one.
10. What would falsify this
Stated as commitments, in decreasing order of how much damage each would do to this paper's claims.
A third Nairobi informal settlement with no documented volunteer-mapping campaign is found to have OpenStreetMap coverage comparable to Kibera's 82.5%. That would directly contradict this paper's proposed mechanism — that campaign presence, not settlement characteristics, explains the gap — and the paper would need to either withdraw the mechanism or substantially narrow its claimed scope.
A re-query of the same two settlements against a later Overture release finds source-composition percentages materially different from those reported in §2, absent an explanable cause (e.g. a new mapping campaign in the interim). That would mean either this paper's own figures are wrong or Overture's merge is less stable over time than assumed, and either would need correcting before the numbers should be relied on.
Independent verification finds Hagen's account of Map Kibera's founding or methodology (§4) is materially inaccurate — for instance, if the mapping was substantially directed by outside researchers rather than resident-led, contrary to the project's own public framing. That would require revising §4's characterization, though it would not touch §2's source-composition measurement.
A methodologically stronger, peer-reviewed study supersedes the Okyere et al. preprint (§6) and finds Google's and Microsoft's Nairobi building data in fact achieves high geometric agreement with reference data. That would weaken §6 and §9's caveat about ML-source reliability, though it would not touch §2's coverage-composition finding, which is a different kind of measurement entirely.
11. Limitations, and what this paper does not claim
Kibera and Mathare are not a representative sample of Nairobi's informal settlements, let alone of informal settlements generally, and §9 states this directly as the paper's most important limitation.
It does not claim ML-detected building footprints are geometrically reliable in Nairobi; §6 reports an independent, if not yet peer-reviewed, finding that they are only moderately so, and treats that as a real caveat rather than an inconvenience to omit.
It does not claim a single, exhaustive account of why Map Kibera succeeded where comparable efforts elsewhere have not; §4 reports what its own cited sources document, not a general theory of successful community mapping campaigns.
It does not claim authoritative settlement boundaries; §2's bounding boxes are approximate and adequate for source-composition ratios, not for authoritative building totals.
And it does not claim that this paper's own evidence settles whether merged datasets are the right choice for every civic-data use case; it establishes that they are a better choice than OpenStreetMap alone for the two specific, checkable cases this paper examined.
12. What would need to be true for this to scale to the next settlement
Kibera and Mathare are two names. Nairobi's informal settlements include Mukuru, Korogocho, Kawangware, Dandora, Huruma and others this series has not individually examined, and every African city of any size has its own list. This paper does not have the resources to run §2's query against all of them, but it is worth naming what a useful next step would actually require, rather than leaving the generalisation as an unexamined assumption.
It would need, at minimum, a documented inventory of which settlements have had a sustained, named volunteer-mapping campaign and which have not — the independent variable this paper's mechanism depends on — cross-referenced against each settlement's current OpenStreetMap-versus-merged-dataset source composition. Kibera and Mathare happen to sit at opposite ends of that independent variable, which is exactly why they made a clean comparison; a systematic version of this paper would need settlements spread across the middle of that range as well, to check whether the relationship is a clean step function (mapped or not) or a more gradual one (partially mapped, intermittently mapped, mapped once and then abandoned).
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 WP27; what we do about the attributes satellites can't see (height, use, condition) is in WP29. A first-person process account is in Building Nairobi a Second Time.
Appendix A — Evidence table
| # | Claim used in this paper | Value | Source |
|---|---|---|---|
| 1 | Kibera buildings, total | 8,130 | Internal pipeline measurement |
| 2 | Kibera buildings from OpenStreetMap | 6,706 (82.5%) | Internal pipeline measurement |
| 3 | Mathare buildings, total | 5,061 | Internal pipeline measurement |
| 4 | Mathare buildings from OpenStreetMap | 794 (15.7%) | Internal pipeline measurement |
| 5 | Sub-Saharan Africa OSM building completeness vs. global mean | 30% vs. 24% | Herfort et al. 20231 |
| 6 | Share of Sub-Saharan African building edits from organised humanitarian mapping | 51.1% | Herfort et al. 20231 |
| 7 | Map Kibera founding month | November 2009 | Map Kibera (mapkibera.org)4 |
| 8 | Google Open Buildings average IoU vs. reference data, Nairobi (preprint, unreviewed) | 0.40 | Okyere, Lu & Brunn 20253 |
| 9 | Microsoft Building Footprints average IoU vs. reference data, Nairobi (preprint, unreviewed) | 0.52 | Okyere, Lu & Brunn 20253 |
| 10 | OpenStreetMap coverage for a third Nairobi settlement with no documented volunteer-mapping campaign | — | — |
| 11 | Peer-reviewed (non-preprint) geometric-quality comparison of building datasets for Nairobi | — | — |
Appendix B — Citation register
CITATION-NEEDED — a third settlement's source composition. Appendix A row 10. §9 states the sample-size limitation directly; a third settlement with no volunteer-mapping history would be the cleanest test of this paper's proposed mechanism.
CITATION-NEEDED — peer-reviewed geometric-quality data for Nairobi building footprints. Appendix A row 11. §6's Nairobi-specific IoU figures come from an arXiv preprint, not yet peer-reviewed; §10 names a peer-reviewed supersession as a falsification condition for the caveat raised there.
Benjamin Herfort, Sven Lautenbach, João Porto de Albuquerque, Jennings Anderson and Alexander Zipf, "A spatio-temporal analysis investigating completeness and inequalities of global urban building data in OpenStreetMap," Nature Communications 14, article 3985 (2023), doi:10.1038/s41467-023-39698-6.↩
Erica Hagen, Open Mapping from the Ground Up: Learning from Map Kibera, Making All Voices Count Research Report (Brighton: Institute of Development Studies, 2017). Read directly via the publisher's summary and report page. Documents Map Kibera's founding (2009, with GroundTruth co-founders Erica Hagen and Mikel Maron providing training and platform support to resident mappers), its eight-year span across digital mapping, citizen media and open data, and concludes that "open, community-based data collection can lead to greater trust, which is sorely lacking in marginalised places."↩
Franz Okyere, Meng Lu and Ansgar Brunn, "Evaluating the Quality of Open Building Datasets for Mapping Urban Inequality: A Comparative Analysis Across 5 Cities," arXiv:2508.12872 (2025). Fetched and read directly. An unreviewed preprint as of this writing; cited for its Nairobi-specific IoU and overlap figures, not as a peer-reviewed finding.↩
Map Kibera, "About" (mapkibera.org, accessed 2026). Fetched directly. States: "Kibera in Nairobi, Kenya, was a blank spot on the map until November 2009, when young Kiberans created the first free and open digital map of their own community."↩
