Methodology
What goes in, and how we check it.
This page explains the ingredients and the evidence at the level a buyer needs. It does not explain how to rebuild the engine. A longer client document is shared under NDA during a trial.
What the product measures
Urban Construction Monitor detects and tracks construction sites across whole cities at monthly cadence from January 2018 to the present. For every site it provides a parcel-level polygon, a calibrated probability that it is real construction, an estimated start date with an end date or ongoing status, an estimated area, and a verification status against the national building register. It reports measured activity. It is monitoring, not forecasting.
Ingredients
Optical and radar satellite data
The European Copernicus Sentinel constellation covers every city on the planet several times a month, in all weather, with a continuous archive since 2018. No imagery is bought. This is what makes whole-city monthly monitoring affordable.
Register labels
National building registers and cadastre, in the Netherlands BAG and the Kadaster, supply building footprints, construction years and permit status. They are used to train and verify, and to snap site outlines to real parcels. They are never treated as the single source of truth, because they lag and miss.
Deep learning
Models score every part of the city every month for construction-like change and output calibrated probabilities. A second model combines the detection evidence with the register's footprint history on that exact site to decide the verified tier. Models are always evaluated on cities they never saw in training.
Human audits
For each audited city a reviewer judges the top-100 sites against multi-year high-resolution aerial imagery, blind to the model's scores. Audits produce the accuracy figures we publish and the labels that improve the next city.
The three layers
- Verified. Sites above the verification threshold or confirmed by a human reviewer. The layer to build on.
- Detected. Every site the satellites picked up, whole city, every month. Measured coverage, not a sample.
- Early signals. Activity in the last twelve months that the register has not caught up with yet. Shown as lower confidence and flagged as such.
How we know it works
The verified tier is independently audited by humans on cities the models had never seen: for each audited city a reviewer judges the top-100 sites against multi-year high-resolution aerial imagery, blind to the model's scores. Site timing is benchmarked against register-dated construction starts. Audit figures are shared during a trial, are measured on audited samples at the date of the audit, and are not warranties.
Known limitations, stated plainly
- Timing is estimated. Start and end dates derive from satellite activity traces and carry a typical error of months. They are robust for market analysis and not for legal or contractual use.
- Recently started sites. Works begun in the last two years have had less time to reach the register, so audited precision for them is somewhat below the headline range. They are clearly flagged as early signals.
- Register dependence. Where a register is sparse or absent, the verified tier is built through expert human review instead, which adds onboarding time per city.
- Minimum size. Reliable detection starts around construction footprints of a few thousand square metres. Single-family infill below that scale is out of scope.
Coverage and onboarding
Live: Amsterdam, Utrecht, Den Haag and Eindhoven, monthly from January 2018. Where a national register exists it powers automated verification. Where none exists, onboarding includes an expert human review round against high-resolution imagery plus city-specific training labels, so a human-verified tier ships there too. Verification is a workflow we bring, not a government dataset we depend on.
Sources and attribution
Satellite data. Contains modified Copernicus Sentinel data 2018-2026.
Maps. Basemap by OpenFreeMap, OpenMapTiles, data (c) OpenStreetMap contributors, Open Database Licence.
Buildings and addresses (Netherlands). Contains data from BAG, Kadaster / PDOK.
Aerial photos (Netherlands). Beeldmateriaal Nederland / PDOK.
Elevation (Netherlands). AHN, open data.
Other trademarks and data belong to their owners. Deliverables contain no personal data: polygons, dates and statistics about construction sites and nothing else.
How to read the data
Urban Construction Monitor estimates construction activity from satellite imagery, public registers and statistical models. It shows likely activity, not legal facts. Start and completion dates are estimates. Site boundaries, classes and counts can be wrong or missing. The verified tier is register-assisted with a sample audited by a human reviewer. It is not an inspection of every site. The data are not a survey, valuation, permit confirmation or credit assessment, and must not be the sole basis for lending, underwriting, legal or planning decisions. Please verify independently.