Property tax base mapping from satellite data
Satellite imagery can count, measure, and roughly value taxable structures even where ground records are absent or decades out of date. This page explains which sensors do what, where the method breaks down, and what a municipality can realistically expect from a first survey.
Sensors
- Airbus Pléiades Neo: 30 cm panchromatic, 70 cm multispectral (6 bands including red-edge and near-infrared). At this resolution individual roof panels, skylights, and rooftop water tanks are visible, supporting material classification and footprint delineation down to structures of roughly 10–15 m² in favourable contrast conditions. Tasking revisit is 1–2 days for a given site.
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral (8 bands), plus 8 SWIR bands at 3.7 m. The SWIR channels discriminate roofing materials that look identical in visible light: asbestos-cement versus painted metal, for instance. Revisit approximately 1 day at mid-latitudes. Archive depth extends to 2014.
- Planet PlanetScope: 3 m multispectral (8-band SuperDove), daily global revisit. Too coarse for individual footprint delineation in dense urban fabric, but valuable for temporal consistency checks: confirming that a structure counted in a Pléiades scene was already present six months earlier, or flagging new construction since the high-resolution tasking date.
- Sentinel-2 MSI: 10 m visible and near-infrared, 20 m red-edge and SWIR, 5-day revisit (with both satellites). Useless for individual building detection in most cities but provides free, cloud-corrected time-series context for neighbourhood-level change detection and for validating roof-material spectral classes derived from commercial imagery.
What a roof tells a tax assessor
A roof is a proxy. Corrugated galvanised iron costs less per square metre to install than fired clay tile; fired clay tile costs less than standing-seam aluminium or a membrane-covered concrete slab with a rooftop plant room. These differences are spectrally measurable. WorldView-3's eight multispectral bands, combined with its SWIR channels, can separate at least four or five broad roofing-material classes with reported overall accuracies in the 80–90 per cent range in published studies on comparable urban scenes. That is not a valuation. It is a stratification: a way of sorting tens of thousands of parcels into rough value tiers before any ground inspector arrives.
The practical output is a ranked list, not a precise assessment. Municipalities that have used satellite-derived roof classification as a first-pass audit typically find that 15–30 per cent of structures in fast-growing peri-urban zones are absent from existing tax rolls entirely. The satellite does not tell you the owner's name or the legal tenure status. It tells you the structure exists and gives you a defensible, reproducible basis for deciding which neighbourhoods to prioritise for field verification.
The resolution problem nobody advertises
At 30–50 cm resolution, a building footprint polygon is straightforward to delineate when structures are detached and shadows are short. The problem is the subdivided plot. A single 200 m² roof outline in a Pléiades image could be one large family home, four single-room rental units sharing a party wall, or a small workshop with a residential annex. The satellite cannot tell you which. This ambiguity is not a failure of the sensor; it is a physical limit of what nadir or near-nadir imagery can observe.
Fusion with OpenStreetMap building data helps but does not resolve the problem. OSM coverage in rapidly urbanising areas is patchy and often lags construction by two to five years. Where OSM polygons exist they can anchor the image segmentation and reduce over-merging of adjacent structures. Where they do not, the analyst must choose a segmentation threshold that will systematically undercount units in dense informal fabric and overcount in low-density areas. Honest programme design acknowledges this and builds in a field-verification sample, typically 5–10 per cent of parcels, to calibrate the statistical bias before results are used for billing.
Spectral classification of roofing materials: what the published record shows
WorldView-3's SWIR bands (1195–2365 nm across eight channels) were specifically designed with material discrimination in mind. Asbestos-cement sheeting, which is both a health concern and a common marker of older, lower-value stock, has a distinct SWIR reflectance profile that separates it from galvanised iron and from painted fibre-cement. Several published studies using WorldView-3 SWIR data over African and South Asian cities have demonstrated separability of five to seven material classes using support-vector machine or random-forest classifiers trained on field-collected spectra.
Pléiades Neo lacks SWIR but its six-band multispectral stack, combined with texture features derived from the 30 cm panchromatic band, can achieve three to four class separability reliably. Painted versus unpainted metal, and concrete versus clay tile, are the most consistent discriminations. Confusion between dark membrane roofing and weathered corrugated iron is a known failure mode at this resolution. Sentinel-2 is not useful for per-roof classification in any urban setting where buildings are smaller than roughly 100 m²; it is included in the workflow only as a temporal anchor.
Fusion workflow and deliverable structure
A practical tax-base mapping programme runs in three stages. First, high-resolution tasking over the municipality produces a cloud-free mosaic, typically requiring two to four collection passes in humid tropical climates to achieve acceptable cloud cover. Second, automated segmentation extracts building footprints and assigns each a roof-material class and a confidence score. Third, PlanetScope time-series is queried against the footprint layer to flag structures that appear to have been built or demolished within the preceding 12–24 months, separating the stable stock from recent change.
The deliverable is a GIS layer: polygon per detected structure, with attributes for estimated footprint area, roof-material class, material-class confidence, and a temporal-stability flag. This feeds directly into a municipality's existing assessment workflow; it does not replace the assessor's judgement or the legal process of formal valuation. Satellize's analytics team has applied analogous stratified-estimation methods in agricultural contexts, including the Kingdom of Tonga crop-estimation programme, and the statistical design principles transfer directly to urban property surveys.
Honest limits and what to do about them
Cloud cover is the most consistent operational problem. Tropical cities can have cloud-free windows of only a few days per month, and a single tasking pass will rarely capture the whole municipality in one go. Mosaicking across multiple dates introduces radiometric inconsistency that degrades material classification accuracy unless careful normalisation is applied. Budget for this.
Tall trees overhanging roofs occlude footprints. In areas with dense canopy cover, building detection rates can fall by 20–40 per cent relative to open terrain. Combining LiDAR-derived canopy height models, where available, with the optical imagery improves this, but LiDAR acquisition adds cost and is rarely available for the municipalities most in need of a tax-base survey. The honest answer is that satellite optical methods work best in arid and semi-arid cities and in tropical cities during dry-season windows. They work least well in dense forest-edge settlements and in cities where informal construction uses materials spectrally similar to the surrounding vegetation. Know your study area before committing to a sensor strategy.
Typical figures
| Best spatial resolution (panchromatic) | 30 cm (Pléiades Neo), 31 cm (WorldView-3) |
| Best spatial resolution (multispectral) | 70 cm (Pléiades Neo), 1.24 m (WorldView-3) |
| SWIR material discrimination bands | WorldView-3 only: 8 SWIR bands, 3.7 m resolution, 1195–2365 nm |
| Temporal consistency sensor | PlanetScope SuperDove, 3 m, daily global revisit |
| Minimum detectable structure footprint | ~10–15 m² under favourable contrast; larger in dense shadow or canopy |
| Tasking revisit (commercial) | 1–2 days (Pléiades Neo and WorldView-3 at most latitudes) |
| Archive depth | WorldView-3 from 2014; Pléiades from 2012; PlanetScope from ~2016 globally |
| Roof-material classification accuracy (published range) | 80–90% overall accuracy for 4–7 classes using WorldView-3 SWIR + VIS in comparable urban studies |
| Delivery format | GeoPackage or Shapefile polygon layer with per-structure attributes; optional GeoTIFF classified raster |
| Cloud-cover constraint | Effective collection requires <20% cloud cover per scene; tropical programmes typically require 2–4 collection passes |
Analytics Satellize can run
| Building footprint inventory | Object-based image analysis (OBIA) segmentation on pan-sharpened commercial imagery, with OSM building data as optional anchor | Polygon GIS layer: one feature per detected structure, with footprint area in m² |
| Roof-material classification | Random-forest or support-vector machine classifier trained on field-collected or library spectra; WorldView-3 SWIR for 5–7 classes, Pléiades Neo VIS/NIR for 3–4 classes | Per-polygon attribute: material class label and confidence score (0–1) |
| Property value tier stratification | Ordinal ranking of parcels by roof-material class and footprint area; no absolute valuation, ranked tiers only | Tabular report: parcel count and estimated floor area by tier, by administrative sub-zone |
| Temporal stability flag | Change detection on PlanetScope monthly composites over the preceding 12–24 months, compared against high-resolution mosaic date | Binary attribute on each footprint polygon: stable stock vs. recently constructed or demolished |
| Missing-from-register estimate | Spatial join of detected footprint layer against existing cadastral or tax-roll polygons; unmatched detections flagged as potential omissions | GIS layer of unmatched structures with roof-material class; summary count by neighbourhood |
| Field-verification sample design | Stratified random sampling across material-class and footprint-area strata, sized to achieve ±5 percentage points precision at 90% confidence on classification accuracy | Prioritised field-visit list with GPS coordinates and predicted class, formatted for mobile data-collection tools |
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.