Building age and construction-era proxy classification from spectral decay
Spectral reflectance shifts in weathered roofing materials, moss colonisation, and surface oxidation allow satellite imagery to estimate relative building age without a cadastral registry. The method is a proxy, not a legal record, and re-roofing breaks the signal.
Sensors
- ASI PRISMA: Hyperspectral imager covering 400–2500 nm in 239 contiguous bands at 30 m spatial resolution. The full SWIR range resolves bitumen oxidation, iron-oxide staining, and carbonate crusting on concrete roofs, which are the primary spectral decay signals. Revisit approximately 29 days at nadir, though tasking can improve this.
- Maxar WorldView-3: Eight multispectral bands at 1.24 m resolution plus eight SWIR bands at 3.7 m. The SWIR bands (1195–2365 nm) distinguish metal corrosion states and asphalt weathering at a spatial scale fine enough to isolate individual rooftop panels. On-demand tasking; no fixed revisit.
- ESA Sentinel-2: Thirteen bands from 443 to 2190 nm at 10–60 m resolution, 5-day revisit at mid-latitudes. Free and globally archived since 2015. Resolution is too coarse to isolate individual rooftops in dense cities, but the archive depth enables multi-temporal spectral trajectory analysis across entire districts, which strengthens age-class separation.
- Airbus Pleiades Neo: Four multispectral bands at 30 cm resolution. No SWIR, so spectral decay discrimination is limited compared with WorldView-3, but the spatial resolution resolves surface texture proxies (moss patterning, staining boundaries) that carry age information independent of spectral band position.
What weathering actually does to a reflectance curve
A freshly laid bitumen-felt roof reflects strongly in the near-infrared and absorbs heavily in the visible, giving a characteristic flat, dark signature. Over five to fifteen years, photo-oxidation converts surface hydrocarbons, raising visible-band reflectance and flattening the SWIR absorption feature centred near 1700 nm. Iron-oxide staining from corroding metal flashings adds a diagnostic red-edge shift. Concrete roofs develop calcium carbonate efflorescence, which brightens the 400–700 nm range. Each of these processes leaves a measurable spectral fingerprint.
Moss and lichen colonisation adds a further layer. Chlorophyll absorption at 670 nm and the red-edge inflection near 700 nm are detectable in PRISMA data and, at coarser scale, in Sentinel-2 band 5 (705 nm). Colonisation density correlates broadly with surface age and porosity, though north-facing aspects and high-rainfall climates accelerate it independently of age. This is one of the honest ambiguities the method carries from the start.
Calibration: why known samples are not optional
Spectral decay indices are relative, not absolute. To assign construction-era classes rather than just rank buildings by apparent age, the classifier must be anchored to a training set of structures with documented build dates. Planning authority records, building permits, or historical aerial photography can supply these. In practice, a sample of 50 to 200 rooftops with verified dates, spread across the target city, is enough to fit a supervised classifier for three to five era classes. Finer class resolution is rarely supportable from spectral data alone.
The calibration breaks in two well-understood ways. First, re-roofing resets the spectral clock. A 1960s building with a 2019 roof will classify as new. Second, rooftop equipment, solar panels, and green roofs mask the substrate signal entirely. Any deployment should report the estimated re-roofed fraction as a confidence qualifier, derived either from permit records or from high-resolution change detection between archive epochs.
Sensor choice is a resolution-versus-chemistry trade-off
PRISMA's 239-band hyperspectral coverage gives the richest chemical discrimination but at 30 m per pixel, which in a dense city means each pixel averages several rooftops. It is best suited to low-density suburban morphology or to city-wide era mapping where individual-building precision is not required. The archive is limited: PRISMA launched in March 2019, so multi-temporal trajectories extend back only to that date.
WorldView-3's SWIR bands at 3.7 m resolve individual rooftops in most urban fabrics and provide enough spectral range to distinguish the main decay classes. The cost is commercial tasking expense and the absence of a free global archive. For a city with 200,000 buildings, full coverage requires multiple acquisition strips and careful mosaicking.
Sentinel-2 is the workhorse for large-area, multi-temporal analysis. Its archive from 2015 onward allows spectral trajectory modelling across nine or more years, which partially compensates for coarse resolution. A building that has been darkening steadily in band 11 (1610 nm) since 2015 is almost certainly older than one that appeared bright in 2022. Temporal gradient is a proxy for age even when instantaneous spectral state is ambiguous.
What the output map can and cannot tell a planner
The deliverable is a raster or vector layer assigning each rooftop polygon (or pixel cluster) to a relative age class: typically pre-1970, 1970–1990, 1990–2010, and post-2010, though the class boundaries are adjusted to match the calibration sample distribution. Accuracy on held-out validation samples, reported in published spectral-decay studies using WorldView and Sentinel data, sits in the range of 60–80 per cent for three-class problems. Four or five classes typically drop accuracy by 10 to 15 percentage points.
For land administration, this is useful as a prioritisation layer, not a legal instrument. It can direct field survey teams to districts likely to contain pre-1970 stock, flag potential unregistered infill construction where new spectral signatures appear without corresponding permit records, or support property tax base estimation by identifying areas where assessed values may not reflect structural age. A planning department that understands the re-roofing caveat will use the output correctly.
Vertical structure is invisible to this method. A building that was extended upward retains its original roof, so the spectral age of the roof is the age of the original structure, which may or may not be relevant. Floor-count estimation requires a separate approach, covered in the sibling page on vertical urban growth.
Practical deployment: what a city government should expect
A typical city-scale run on Sentinel-2 data, covering 500 to 1,000 square kilometres, takes two to four weeks from data acquisition to classified output, assuming a calibration sample is already assembled. WorldView-3 coverage of the same area requires scheduling commercial tasking and is better scoped as a targeted survey of specific districts rather than a wall-to-wall acquisition.
Satellize structures this kind of analysis as a calibrated proxy product: the output includes the era-class raster, a per-polygon confidence score derived from spectral distance to class centroids, and a written methodology note that documents the calibration sample, the re-roofed fraction estimate, and the specific spectral indices used. The Tonga crop-estimation programme established the internal workflow for proxy classification with honest uncertainty reporting, and the same discipline applies here.
Updating the analysis annually using Sentinel-2 costs a fraction of the initial run. New construction appears as bright, spectrally young clusters; demolition and redevelopment show as abrupt spectral resets. A difference layer between annual outputs is often more actionable for planning enforcement than the absolute age map.
Typical figures
| Spatial resolution (hyperspectral) | 30 m (PRISMA); individual rooftop isolation requires supplementary footprint data |
| Spatial resolution (multispectral SWIR) | 3.7 m SWIR / 1.24 m panchromatic (WorldView-3); resolves individual rooftops in most urban morphologies |
| Spatial resolution (free archive) | 10 m (Sentinel-2 visible/NIR), 20 m (SWIR bands 11 and 12); pixel typically covers multiple rooftops in dense areas |
| Revisit | 5 days at mid-latitudes (Sentinel-2 combined A+B); ~29 days nadir (PRISMA); on-demand (WorldView-3, Pleiades Neo) |
| Key spectral bands for decay detection | SWIR 1610 nm and 2190 nm (bitumen oxidation, carbonate); red-edge 705 nm (moss/lichen); visible 670 nm (chlorophyll absorption) |
| Archive depth | Sentinel-2: from 2015; PRISMA: from March 2019; WorldView-3: from 2014 (commercial tasking archive) |
| Minimum era classes reliably separable | 3 classes at 60–80% accuracy (published studies); 4–5 classes reduce accuracy by ~10–15 percentage points |
| Calibration sample requirement | 50–200 rooftops with verified construction dates, distributed across the study area |
| Key accuracy caveat | Re-roofed buildings misclassify as newer stock; solar panels and green roofs mask substrate signal |
| Delivery formats | GeoTIFF raster (era-class + confidence band), GeoPackage vector polygons, PDF methodology note |
Analytics Satellize can run
| City-wide era-class raster | Supervised spectral classification (Random Forest or SVM) on SWIR oxidation and red-edge moss indices, calibrated against permit-verified training samples | GeoTIFF with era-class values and per-pixel confidence score; accompanying methodology PDF |
| District-level age distribution summary | Zonal statistics aggregated over planning zones or administrative boundaries | Tabular report and choropleth map layer showing proportion of pre-1970, 1970–1990, 1990–2010, and post-2010 stock per district |
| Spectral trajectory analysis (multi-year) | Annual median composites from Sentinel-2 band 11 and band 12 time series; linear trend fitting per pixel cluster | GIS layer flagging pixels with statistically significant brightening or darkening trends, indicating new construction or accelerated decay |
| Re-roofing fraction estimate | Change detection between two or more WorldView-3 or Pleiades Neo acquisitions; abrupt spectral resets on stable footprints flagged as probable re-roofing events | Confidence qualifier layer appended to era-class output; district-level re-roofing rate table |
| Unregistered infill detection | Comparison of spectrally young rooftop clusters appearing after a reference date against building permit records supplied by the client | Vector point layer of candidate unregistered structures for field verification; ranked by spectral confidence |
| Property age proxy for tax base analysis | Era-class output joined to cadastral parcel boundaries; age-class distribution summarised per parcel where footprints allow | Parcel-level GeoPackage with era-class attribute for integration into assessment workflows |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.