Height is the dimension every 3D city model needs and almost none of the open data actually provides. This paper documents the gap as we found it in Nairobi specifically, and the heuristic built to stand in for it while the real fix — a proper remote-sensing height dataset — waits on infrastructure that doesn't exist yet.
The four tiers, drawn honestly: the fainter the building, the less we actually know about its height.
1. What this heuristic is honest about, and what it isn't
This paper makes a narrower claim than it might look like it makes, worth stating precisely before defending it.
What it argues from. A direct measurement of Overture's own data sparsity (4,203 of 1,277,511 buildings carry a recorded height), a described four-tier estimator whose logic is fully specified below, and the resulting distribution, reported without smoothing. These are reproducible: anyone with the same Overture release and the same pipeline gets the same table.
What it argues from, more cautiously. The premises behind Tiers H and m — that informal-settlement buildings skew low-rise, and that height decays with distance from a commercial core — are drawn from published urban-geography and informal-settlement literature, not independently validated against measured Nairobi heights. §6 below reports a literature search specific to this paper's own premises that materially complicates the first one, and the paper engages that complication directly rather than holding the premise unchanged.
What it does not argue from. It does not argue that the heuristic is accurate; it argues that the heuristic is honest about its own confidence, which is a different and more modest claim. It does not re-derive the later satellite-measured result, which supersedes Tiers m and l with actual observation once Earth Engine access existed; this paper is a snapshot of the interim state, not a claim that the interim state stayed correct.
What would refute it. Given in full in §9. The sharpest version: if the distance-and-land-use premise in Tier m is shown, on direct comparison against the satellite-measured heights now available, to perform no better than a flat per-building constant, the paper's central design decision — that a structured guess beats an undifferentiated one — would be wrong, and the honest response would be to say so rather than keep defending the tiering as a good idea in principle.
2. How sparse "sparse" actually is
Overture Maps' buildings layer carries height and num_floors fields wherever a contributing source (almost always OpenStreetMap, in practice) happened to tag them. Across the full 1,277,511-building Nairobi metro extract, 4,203 buildings — 0.3% — have either field populated. In a small test extract centered on the CBD, where OSM tagging density is highest anywhere in the city, the figure was still only around 1%. Outside the CBD, height data is close to absent.
This is not an OpenStreetMap failure specifically — height is a field a human mapper has to notice, measure or estimate, and manually enter, for every single building, and almost nobody does that work at scale unless there's a specific reason to (a landmark, a permit dispute, a survey). It's a structural gap in volunteer geographic data generally, not a Nairobi-specific or OSM-specific shortcoming.
3. The real fix, deferred on purpose
A better source exists and is specifically validated for this city: Google's Open Buildings Temporal dataset provides building presence, height, and fractional building counts as a ~4-metre-resolution raster spanning 2016–2023, licensed CC-BY/ODbL for free commercial use. Independent, peer-reviewed research evaluating height-prediction accuracy across three cities — Nairobi, Kathmandu and Quito — against ICESat-2 laser-altimetry ground truth found Open Buildings Temporal achieved a 2.49-metre mean absolute error specifically for Nairobi1 — accurate enough to distinguish a bungalow from a three-story building reliably, which is the resolution a usable city model actually needs.
We did not pull that dataset at the time this heuristic was built. It required a Google Earth Engine project tied to a real Google Cloud account, which hadn't been provisioned when this pipeline was built, and setting one up wasn't that session's job. The decision — made explicitly, not by default — was to estimate heights with visible, honest confidence levels, and replace the lower-confidence estimates with real Open Buildings Temporal data once that access existed, rather than either blocking the entire buildings pipeline on an unrelated infrastructure dependency or shipping a height field that looks measured when it isn't. The access was later provisioned, and a later paper reports what happened when the real raster finally arrived: a genuinely surprising result, since it disagreed with Overture's own surveyed-height attribute by 6.84 metres on average — worse than its validated 2.49-metre error against independent laser altimetry — implying the "surveyed" field this heuristic treats as the gold-standard Tier M was, for a meaningful share of buildings, probably the less trustworthy of the two sources all along.
That finding does not change anything reported in §5 below, which documents the heuristic's distribution at the time it was estimating 99.7% of the city. But it is worth flagging here, early, because it means Tier M's status as "the measured tier, checked first" rested on an assumption — surveyed beats estimated — that turned out not to hold up against independent validation. The heuristic was honest about the confidence of Tiers H, m and l. It was not equally skeptical of Tier M, and that is a real asymmetry this paper should have caught sooner than it did.
4. The heuristic: four tiers, checked in order
For every building without a direct Overture height/floor value, the estimator checks, in order:
Tier H — informal settlement. If the building's centroid falls inside one of six hand-drawn approximate bounding boxes for Nairobi's major informal settlements (Kibera, Mathare, Mukuru, Korogocho, Kawangware, Dandora), height is estimated at 3.5 metres — one story. This was grounded, at the time, in a general pattern commonly cited for Nairobi's informal housing stock: that it skews heavily toward single-storey shack construction. §6 reports what a focused literature search specific to this premise actually found, and it complicates this tier substantially.
Tier m — distance and land-use. For everything else, we compute distance from the same KICC/Times Tower midpoint anchor the whole pipeline uses, bucket it into bands (8 floors within 300m of the core, 4 floors to 800m, 2 floors to 2km, 1.5 floors to 6km, 1 floor beyond), and apply a multiplier from Overture's land-use classification where one intersects the building (industrial and similar classes stay near 1.0 — tall single-volume structures, not many floors; commercial areas get a 1.2–1.3x bump; hospitals 1.4x; farmland-adjacent structures scale down). This rests on a standard premise in urban economics — that building height tends to decay with distance from a city's commercial core, because land value does and vertical construction is the market's response to expensive land2 — applied here as a reasonable starting prior, not as something independently validated against real Nairobi height measurements at the time it was built.
Tier l — fallback. Where no land-use polygon intersects the building at all, the same distance-band floor count applies with no land-use multiplier. This is the least-informed tier and, as the results below show, by far the largest.
Tier M — measured, checked first, before any of the above: Overture's own height or floor-count value, converted at 3.2 metres per floor where only a floor count is given. §3 already flagged the problem with treating this tier as simply correct.
5. The distribution, reported without smoothing it over
| Confidence | Buildings | Share |
|---|---|---|
| M — measured | 4,203 | 0.3% |
| H — informal settlement | 87,459 | 6.8% |
| m — distance + land-use | 246,790 | 19.3% |
| l — fallback | 939,059 | 73.5% |
Most honest sentence in this series: for three-quarters of Nairobi's buildings, our height is a reasonable prior, not a fact — and the map says so, per building, in color.
The dominant tier is the least-informed one, by a wide margin. That's not a flattering number to publish, and it would have been easy to leave it out or bury it in a methods appendix nobody reads. We're reporting it as the headline finding instead, because the alternative — a height column that looks the same regardless of whether it's measured, well-reasoned, or a bare guess — is a worse kind of honesty failure than an unflattering distribution. The land-use layer that tiers m and l split on is itself sparse (3,782 features across roughly 1,600 square kilometers), which is the direct, traceable reason so much of the metro falls into the flat fallback: most of the city simply has no Overture land-use polygon to key off at all.
6. What a focused search found about Tier H's premise
Tier H rests on a specific claim: that Nairobi's informal-settlement housing stock is overwhelmingly single-storey. I went looking for the literature behind that claim rather than continue to treat it as settled background knowledge, and what I found is more heterogeneous than the original heuristic assumed.
A 2020 IIED working paper on Nairobi's housing markets, based on fieldwork, focus groups and a review of Kenya's 2019 census and prior surveys, draws a sharper distinction than "informal settlement equals single storey."3 Citywide, 36% of Nairobi's households live in single-room tenements — purpose-built rental blocks, typically multi-storey — while a separate 34% live in single rooms made of iron sheets or mud, the shack typology Tier H was actually modelling, concentrated in settlements like Kibera, Mathare and Mukuru.3 Tenements are not a marginal variant: the same paper reports they average seven walk-up floors and sometimes reach ten, with Pipeline given as a named example, and that single-room tenement housing is in fact Nairobi's most prevalent shelter type overall, ahead of the shack typology.3
Crucially, the two typologies do not sort cleanly by settlement; they sort by sub-area and by land-tenure history within a single settlement. For Mukuru specifically, the paper reports that roughly 90% of residents are tenants in single- or double-storey metal shacks, with a 2016 profile finding 30% of shacks had become double-storeyed and the settlement still having "few tenements" — closer to Tier H's assumption, though not exactly it, since a flat one-storey estimate still undercounts the double-storey third.3 For Mathare, the picture is explicitly uneven within the settlement: villages with secure, de facto land tenure (the paper names Kosovo and Gitathuru) are "increasingly transitioning to tenements," driven by rising demand from migrants and students, while villages with contested or public-land ownership (Kiamutisya is named) remain shack-dominated.3 The same settlement, on this account, contains both a sub-area where Tier H's one-storey assumption is roughly right and a sub-area where it is systematically wrong by a factor that can reach seven or ten.
This matters for the heuristic in a specific, falsifiable way. Tier H draws one hand-drawn bounding box per settlement and applies one height to everything inside it. The literature says the real variation is at a finer grain than the settlement — it is the village or even the plot, keyed to land-tenure status and proximity to specific demand drivers (a college, a main road), not to which of six named slums a building happens to sit in. A bounding box drawn around all of Mathare will therefore flatten Kosovo's emerging tenements into the same one-storey guess as Kiamutisya's shacks, understating height for exactly the buildings where the understatement is largest.
7. What this means for the reported numbers in §5
What this literature search does and does not change is worth stating precisely. It does not change the 87,459-building, 6.8% figure in §5's table — that is what the heuristic, as actually built, produced, and it is reported honestly regardless of whether the heuristic's premise holds up. What it changes is how much confidence Tier H's label deserves relative to Tiers m and l. The original design ranked H above m and l specifically because a settlement-specific, literature-grounded assumption was presumed more reliable than a generic distance-decay prior. §6's finding complicates that ranking: Tier H's assumption is accurate for some fraction of its 87,459 buildings and specifically wrong, by a large margin, for buildings inside the minority of informal-settlement sub-areas that have transitioned to tenements. I do not have a verified estimate of what fraction that is citywide, and §9's citation register says so rather than guessing a number to fill the gap.
The honest fix is not to silently reclassify Tier H buildings into a different confidence bucket without evidence for where the boundary actually falls. It is to flag, here, in the paper, that Tier H's confidence label should be read as "a literature-grounded guess, now known to be heterogeneous in a specific, documented way" rather than "the most reliable of the three estimated tiers," which is the ranking its position above Tiers m and l implies.
8. Why this is built to be replaced, not kept
Every estimate carries its confidence code alongside it in the tile data, all the way to the rendered map (/nairobi/map/, where dot color encodes confidence directly). That was a deliberate design constraint, not a nice-to-have: when real Open Buildings Temporal data eventually gets pulled, the refinement is a per-tile, per-confidence-tier replacement — swap l and m estimates for measured raster values where the raster has coverage, leave M alone pending the re-assessment §3 now flags as necessary, reassess H against the same raster to check whether the single-storey assumption actually holds in each specific sub-area §6 describes as heterogeneous. None of that requires re-deriving footprint geometry, re-running the tiling pipeline, or touching the parts of the dataset that are already solid. The estimate was always meant to be provisional. Labeling it that way, tier by tier, is what makes the provisional part actually true instead of just claimed — and the later paper confirms the replacement did happen, with results this paper's own Tier M did not anticipate.
9. Objections I take seriously
"If Tier H's premise is this heterogeneous, the paper should have re-estimated it, not just flagged it." I considered this and decided against it, for a reason worth stating rather than hiding behind. Re-estimating Tier H credibly would require either a settlement-by-settlement, sub-village-by-sub-village land-tenure survey — which §6's own source shows varies at a finer grain than any bounding box this pipeline currently draws — or the satellite-measured heights the later paper eventually obtained, which supersede the heuristic entirely rather than patch it. Adjusting Tier H's single number without either of those would replace one unvalidated guess with another unvalidated guess, dressed in more confident language. §7 states plainly that the honest move was disclosure, not a cosmetic correction.
"A heuristic this wrong about its own best-grounded tier undermines trust in the whole confidence-tier scheme, including Tiers m and l." This is fair and I do not think it fully resolves. What distinguishes Tier H's problem from a reason to distrust the whole scheme is that §6 found a specific, documented heterogeneity — tenure status and proximity to named demand drivers — not a general reason to doubt structured estimation as such. Tiers m and l rest on a different, more general premise (distance-decay of height with distance from a commercial core) that has its own literature and its own objection, addressed next.
"The distance-decay premise behind Tier m assumes an undistorted land market, but Nairobi's informal tenement construction documented in §6 happens largely outside formal planning control — which is exactly the condition under which the 'natural' density gradient a market-based premise predicts can invert." This is a serious point. Bertaud's own account of density gradients notes that in cities where planning regulation (height caps near the core, looser limits at the periphery) distorts the market response to land value, high-rise construction can appear on the periphery and low-rise near the core — the inverse of the simple gradient.2 Nairobi's informal tenement construction, built largely outside formal planning and height-regulation control altogether, is not quite either case Bertaud describes, but it is a third kind of distortion the simple distance-decay premise does not model: height driven by local demand and land-tenure security rather than distance from the CBD at all. Tier m does not currently account for this, and that is a real limitation rather than a solved problem.
"Flux has a commercial interest in this conclusion." Flux builds the Nairobi digital twin this heuristic is part of, and a plausible-looking height field is a direct component of the product's polish. The conflict is real. The mitigation is what this paper actually does: §5 reports the unflattering 73.5%-fallback number as the headline rather than a footnote, and §6 reports a literature search that complicates the paper's own best-grounded tier rather than one that only confirms it. A reader does not have to trust Flux's motives to check that the reported distribution matches what the described pipeline produces, or that the cited sources say what this paper says they say.
10. What would falsify this
Stated as commitments, in decreasing order of how much damage each would do to this paper's claims.
A direct comparison of Tier H's estimates against the later satellite-measured heights, restricted to the six named informal settlements, finds no meaningful difference in error between Tier H buildings and Tier m/l buildings. That would falsify this paper's central claim that a literature-grounded, settlement-specific prior is better-informed than a generic distance-decay prior — the two tiers would be equally uninformed in practice, regardless of which published premise each drew on.
A sub-village-level survey of one or more of the six named settlements finds the tenement/shack split §6 describes for Mukuru and Mathare does not generalize — for instance, finds Kibera, Korogocho, Kawangware or Dandora are in fact close to uniformly single-storey. That would weaken §6's complication of Tier H specifically for the settlements not directly covered by the IIED paper's case studies, which named only Mathare and Mukuru in the detail this paper draws on.
The distance-decay premise behind Tier m, tested directly against measured heights, performs no better than a flat per-building constant with no distance or land-use information at all. That would falsify this paper's §1 commitment that a structured guess beats an undifferentiated one, and is the test named in §1 as the sharpest available.
A land-tenure dataset for Nairobi's informal settlements becomes available at a finer grain than this pipeline's six hand-drawn bounding boxes, and shows the tenure-tenement correlation §6 reports does not hold outside the two settlements it was documented for. That would weaken, though not eliminate, the specific mechanism §6 proposes for why Tier H's error is concentrated in certain sub-areas rather than spread evenly.
11. Limitations, and what this paper does not claim
The four-tier heuristic is not claimed to be accurate; it is claimed to be honestly labelled, and the two are not the same claim. §5's headline number is the clearest statement of this distinction.
It does not claim Tier H's single-storey assumption is wrong everywhere it is applied; §6 and §7 both state the finding is that the assumption is heterogeneous within settlements, not uniformly false.
It does not claim a verified estimate of what share of Tier H's 87,459 buildings fall inside a tenement-transitioning sub-area rather than a shack-dominated one; Appendix B registers this as an open measurement rather than an estimated one.
It does not claim the distance-decay premise behind Tier m has been validated against measured Nairobi heights; §9's third objection states the specific reason Nairobi's informal construction may not fit that premise's assumptions.
And it does not claim Tier M's status as "the measured tier" was ever sound; §3 states plainly that this paper's heuristic trusted Tier M more than later evidence, reported in a companion paper, turned out to justify.
12. Data & methods availability
Input data and bbox as in the pipeline paper (Overture 2026-06-17.0). The estimator is the build_nairobi_tiles command in the Flux repository; the four tiers, the distance bands (8 floors ≤300 m from the KICC/Times Tower midpoint, decaying to 1 beyond 6 km, at 3.2 m per floor), the land-use multipliers, and the six settlement bounding boxes are all in that one file, so the full distribution table reproduces from public data with one command. The distance-decay premise follows the standard urban-economics density-gradient literature;2 it has not been independently validated against measured Nairobi heights at the pipeline level — a companion paper reports what the later satellite-measured validation found for the dataset as a whole.
Companion papers
How the buildings themselves were obtained and tiled: WP27. Why single-source map data systematically misses the settlements this paper's H-tier covers: WP28. What happened when the real satellite-measured heights arrived: WP32. A first-person process account: Building Nairobi a Second Time.
Appendix A — Evidence table
| # | Claim used in this paper | Value | Source |
|---|---|---|---|
| 1 | Buildings with a direct Overture height/floor value | 4,203 of 1,277,511 (0.3%) | Internal pipeline measurement |
| 2 | Tier distribution: M / H / m / l | 0.3% / 6.8% / 19.3% / 73.5% | Internal pipeline measurement |
| 3 | Open Buildings Temporal MAE vs. ICESat-2 laser altimetry, Nairobi | 2.49 m | Watson & Elliott 20251 |
| 4 | Open Buildings Temporal MAE vs. Overture's own surveyed attribute, Nairobi (companion measurement) | 6.84 m | WP32 |
| 5 | Share of Nairobi households in single-room tenements (citywide) | 36% | Mwau, Sverdlik & Makau 2020, citing Mutinda et al. 20193 |
| 6 | Share of Nairobi households in single-room iron-sheet/mud shacks (citywide) | 34% | Mwau, Sverdlik & Makau 2020, citing Mutinda et al. 20193 |
| 7 | Average floors per tenement building; maximum cited (Pipeline) | ~7 average; up to 10 | Mwau, Sverdlik & Makau 20203 |
| 8 | Mukuru: share of slum shacks that had become double-storeyed by 2016 | 30% | Mwau, Sverdlik & Makau 2020, citing Lines & Makau 20173 |
| 9 | Mukuru: share of residents who are tenants | ~90% | Mwau, Sverdlik & Makau 20203 |
| 10 | Share of Tier H's 87,459 buildings located in a tenement-transitioning (rather than shack-dominated) sub-area | — | — |
| 11 | Validated performance of the distance-decay premise (Tier m) against measured Nairobi heights | — | — |
| 12 | Land-tenure data for Nairobi's informal settlements at finer-than-bounding-box grain | — | — |
Appendix B — Citation register
CITATION-NEEDED — share of Tier H buildings in tenement-transitioning sub-areas. Appendix A row 10. §6 and §7 document that the heterogeneity exists and name the mechanism (land tenure, proximity to specific demand drivers); no source quantifies it citywide or for the four of six named settlements not covered by the IIED paper's case studies (Kibera, Korogocho, Kawangware, Dandora).
CITATION-NEEDED — validated accuracy of the distance-decay premise. Appendix A row 11. Tier m's premise is grounded in general urban-economics literature,2 not validated against measured Nairobi heights at the time this heuristic was built; §10 names the direct test that would settle it.
CITATION-NEEDED — fine-grained land-tenure data for informal settlements. Appendix A row 12. §6's finding that tenure status predicts tenement-vs-shack transition is drawn from two settlements' worth of case-study detail in one source; a land-tenure dataset at village or plot grain, if one existed, would let Tier H's bounding boxes be subdivided rather than flattened.
C. Scott Watson and John R. Elliott, "Narrowing the gap for city building height predictions," Scientific Reports 15, article 29913 (2025), doi:10.1038/s41598-025-15929-2.↩
Alain Bertaud, Order without Design: How Markets Shape Cities (Cambridge, MA: MIT Press, 2018). Formalises the standard urban-economics premise that building height responds to land value, which typically declines with distance from a city's commercial core — and separately documents that planning regulation (e.g. height caps near the core, looser limits at the periphery) can invert the "natural" gradient, producing high-rise construction at the periphery and low-rise near the core.↩
Baraka Mwau, Alice Sverdlik and Jack Makau, Urban Transformation and the Politics of Shelter: Understanding Nairobi's Housing Markets (London: International Institute for Environment and Development, 2020), ISBN 978-1-78431-777-5, http://pubs.iied.org/10876IIED. Fetched and read directly (sections 4.5–4.7 and 5.1–5.2). Reports, citing KNBS (2018a) and Mutinda et al. (2019), that 36% of Nairobi households live in single-room tenements (averaging seven floors, up to ten) and 34% in single-room iron-sheet/mud shacks; and, citing Lines and Makau (2017) and the authors' own Mathare fieldwork, that Mukuru remains roughly 90%-tenant single/double-storey shacks with 30% of shacks double-storeyed by 2016 and "currently few tenements," while Mathare's villages split by land-tenure status between tenement-transitioning (Kosovo, Gitathuru) and shack-dominated (Kiamutisya).↩
