Building stock, measured · ALCBT geostock briefing

Bottom line

Three findings that should change how the MRV baselines are used

Thailand is modeled roughly two times too high

The seeded baseline claims 18.6 billion m² of floor area. Satellite-derived stock finds 9.5 billion m², including structures the model does not even intend to cover. The 15 billion m² residential figure cannot be reproduced from measured data.

Indonesia, Vietnam and Cambodia are modeled far too low

Measured stock exceeds the modeled baselines by a factor of 3.2 to 4.8. Part of this is scope (all structures versus occupied buildings), but the direction is unambiguous: the current baselines understate the building stock.

India quantifies the uncertainty band

Two independent measurements bracket the truth: the quality-filtered GOBS dataset is 40 percent above the baseline, the unfiltered Overture dataset 112 percent above. The same band applies, mentally, to the other four countries.

Findings

Measured and modeled built-up area disagree in both directions, in billion m²

Indonesia209 M buildings measured

28.2 measured

5.9 modeled

+377%gap

India · GOBS189 M buildings, quality-filtered

27.6 measured

19.7 modeled

+40%gap

India · Overture400 M buildings, unfiltered

41.8 measured

19.7 modeled

+112%gap

Vietnam80 M buildings measured

11.1 measured

2.6 modeled

+321%gap

Thailand53 M buildings measured

9.8 measured

18.6 modeled

−47%gap

Cambodia8.7 M buildings measured

1.02

0.32

+217%gap

Measured · satellite-derived, all structures Modeled · seeded baseline, res + com + inst floor area

Proof

The Thailand anomaly and the India validation carry the argument

Thailand · the baseline cannot be reproduced

6.7 B m²measured footprint

9.8 B m²measured built-up (EE heights, 13 Jun)

18.6 B m²modeled baseline

Thailand has 5 percent of India's population but is seeded with 94 percent of India's floor area. Even counting every detected structure of any type, measured stock reaches barely half the modeled value. The 15.0 B m² residential input (source: HEAT GmbH / IEA 2023 estimate) is the dominant error candidate and should be reconciled first.

India · two instruments, one consistent picture

−2.5%floor-factor deviation, GHSL vs GOBS

4.21 / 4.51 mmean building height, GHSL vs GOBS

0.0state-sum error vs national totals

The GHSL-derived floor factor reproduces the factor implied by AEEE's independent per-building GOBS dataset within 2.5 percent, and mean heights agree within 7 percent. The method transfers: the same pipeline produced the Indonesia, Thailand, Vietnam and Cambodia values.

Method

Three open datasets feed one comparable schema with full provenance

Measured sources

GOBS v1.0 (AEEE) · India, per building, 2023
Overture Maps 2026-05 · Google + Microsoft + OSM footprints
GHSL R2023A · built volume and surface, 1 km

→

geostock schema

sources · registry with data-quality class
admin_stock · country x admin unit x land use x year
row-level security · public read, service-role write

→

Comparison view

v_measured_vs_modeled · national gap per source
new measured rows light up automatically
queryable now via REST bridge views

✓

Verification status. India state sums reproduce GOBS national totals exactly (189.24 M buildings, 19.22 B m² footprint, 27.64 B m² built-up). All seeds are idempotent and committed; raw extracts are versioned in the repository. Earth Engine admin-1 height refinement was executed on 13 June 2026; per-building confidence filtering remains the next quality step.

Limits of interpretation

Upward gaps are upper bounds: scope and filtering differ between measured and modeled

Scope

Measured counts every structure, the model does not

Satellite footprints include industrial buildings, sheds, informal and derelict structures. The modeled baseline covers occupied residential, commercial and institutional floor area. Measured exceeding modeled is partly definitional.

Filtering

The India dual measurement calibrates the inflation

Unfiltered Overture exceeds quality-filtered GOBS by a factor of about 1.5 on built-up area. Discount the Indonesia, Vietnam and Cambodia gaps accordingly before quoting them; the direction of the finding survives the discount.

Heights

Heights now measured per province

Since 13 June, built-up area uses admin-1 mean heights from Open Buildings Temporal (Earth Engine) for matched provinces (Thailand 69/77, Cambodia 20/25, Vietnam 27/34, Indonesia 9/38); unmatched units use the measured national mean.

Vintage

Measurement years differ from the baseline year

GOBS reflects 2023, Overture 2026, the modeled baseline 2025. At observed stock-growth rates this explains only a few percent of the gaps, but consistent vintage alignment belongs in the reconciliation.

Recommended actions

Re-baseline before scenario results are quoted externally

Reconcile Thailand first, then adopt measured floor areas where they survive scrutiny

Phase 0 write-up per country: decompose each gap into scope, filtering and genuine baseline error, then decide per country to re-baseline, keep with documentation, or blend. Thailand is the priority because the error direction makes current emissions estimates overstated.

Maxine · Dietram sign-offanalysis exists, decision pending

Surface provenance in the dashboard so users see measured versus modeled

Phase 2 frontend: source toggle in the emissions pillar, 36-state India view, methodology section with these caveats. Branch with Cloudflare preview; production numbers unchanged until sign-off.

Maxinebranch feat/gobs-integration

Completed 13 June: Earth Engine admin-1 height refinement

Province-level measured heights now drive the built-up calculation for the four Southeast Asian countries. Remaining quality step: per-building confidence filtering (GOBS-equivalent) and the land-use split via AEEE.

Donerefined values live in geostock