Neighbourhood classification
A typology of small areas estimated from census-grade statistics and held constant across countries, so a class means the same thing in Lyon as in Leipzig.
Geospatial data · Denmark
Location, measured.
We produce small-area estimates of how people live — derived from official statistics, multispectral satellite imagery and the composition of places, and expressed in a classification that is identical across countries. The unit of analysis is the area, never the individual.
01
What we do
We turn heterogeneous spatial evidence into structured, comparable datasets that other companies build on and bring to market under their own names.
Remote sensing
Self-supervised representation learning and semantic segmentation on multispectral and street-level imagery recover what statistics do not record: built form, building height and era, vegetation and change over time.
Audience estimation
Each audience is a vector of indices over the classification (100 = national mean), estimated from official covariates and country-specific modelled survey responses, then propagated to every postal area. A brief-to-audience engine exposes the same machinery in plain language.
Activation-ready delivery
Versioned files and feeds at postal-code or hexagonal-grid resolution, structured for programmatic platforms, data marketplaces and our partners' own products.
02
Six layers, one score per area
Each layer is an independent, partial observation of a place. We fuse them at area level into a single estimate whose provenance can be traced layer by layer.
Official statistics
Income, population, housing and household data from national statistics offices — the ground truth every model is calibrated against.
Place fingerprint
What a neighbourhood contains — shops, services, amenities, transport. A profile built from places, not people.
Satellite imagery
Machine-learned features from multispectral imagery: built form, building footprints and heights, vegetation and change over time.
Street-level analysis
Building era, type and condition read from façades, where imagery and regulation allow.
Classification
A neighbourhood typology that gives every area a comparable identity across countries.
Modelled survey
Attitudes, interests and purchase intent modelled per country — run fresh for each market, never copied across borders.
Every layer is joined at area level — postal codes and hexagonal grid cells. No personal identifiers, cookies or device IDs enter the pipeline.
03
How we work
- Step 1
Measure
Collect official statistics, imagery and place data for every area in a market.
- Step 2
Classify
Assign each area to the typology. Models are trained in markets where the reference answer is known, then transferred.
- Step 3
Score
Estimate every audience per class against the national mean (100) and validate out of sample on held-out areas and official data.
- Step 4
Deliver
Ship the taxonomy, the index and area-level files, ready to activate — and refresh them as new data arrives.
04
Selected methods
The toolkit behind the classification and the audiences. None of it requires data about individuals.
- Small-area estimation
- Census and registry covariates disaggregated to postal and grid level, with uncertainty carried through rather than discarded.
- Hierarchical hexagonal indexing
- A discrete global grid of hexagons at nested resolutions as the common spatial key — equal-area neighbours, no administrative boundary artefacts.
- Self-supervised geospatial embeddings
- Representations learned from imagery and place composition without labels, then fine-tuned against official ground truth.
- Spectral indices and segmentation
- Vegetation and built-up indices from cloud-masked seasonal composites; building footprints, heights and roof forms from segmentation models.
- Modelled survey responses
- Attitudinal and purchase-intent measures from synthetic respondents scored by semantic similarity, generated separately for each country and benchmarked against published survey data.
- Out-of-sample validation
- Every model is trained where the truth is known and tested on held-out areas and markets before release; audiences that cannot be validated are not shipped.
05
Principles
Areas, not people
We describe places. Nothing we build identifies, follows or profiles an individual.
Statistics teach the model
Official data is the reference every model learns from and is tested against. Commercial panels are a benchmark, never the truth.
Every number reproducible
Each audience has a full index row, and every area-level file can be rebuilt from it.
Built to travel
One framework across markets, with local data and local modelling in each country.
06
What partners receive
- Units
- Postal codes and hexagonal grid cells. Finer grids are internal building blocks, never delivered targeting units.
- Files
- A taxonomy of audiences, an index file (100 = national average) and an area file flagging membership with household and population counts.
- Activation
- Formatted for programmatic platforms, data marketplaces and identity partners that accept geographic input.
- Refresh
- Re-scored when new official data or imagery arrives, with versioned releases.
07
About
Rainbow Insights is a Denmark-based geospatial data house. We work as a research and data partner to companies in Europe and beyond, building the classifications, models and datasets they bring to market.
The company is research-led. Most of what we deliver began as a methodological question: how to measure a place more faithfully, more cheaply or more consistently than the classifications that came before.
08
Get in touch
Tell us what you are building and which markets you need. We reply within two working days.
- Based in
- Denmark