Urban surface material classification using spectral indices
Multispectral and shortwave-infrared reflectance ratios separate roofing and paving materials that visible bands cannot distinguish, supporting heat-island modelling, runoff coefficients and informal-settlement characterisation.
Sensors
- Sentinel-2 MSI: 13 spectral bands including SWIR at 1610 nm and 2190 nm, 10 m resolution in visible/NIR and 20 m in SWIR, 5-day revisit at the equator. Free archive from 2015. The 20 m SWIR bands are the primary discrimination tool for surface materials but require resampling to match the 10 m visible bands.
- Landsat 8/9 OLI: SWIR-1 (1570 nm) and SWIR-2 (2110 nm) at 30 m resolution, 16-day revisit per satellite (8-day combined). The 30 m pixel is coarser than Sentinel-2 and causes significant mixed-pixel problems in dense urban fabric, but the archive extends to 1972 (Landsat 1 MSS) and the OLI radiometric calibration is well-documented and stable.
- Maxar WorldView-3 SWIR: 8 SWIR bands between 1195 nm and 2365 nm at 3.7 m resolution, plus 8 multispectral bands at 1.24 m. The only commercially available sub-5 m SWIR sensor. Resolves individual roof panels and narrow paved lanes that are sub-pixel noise in Sentinel-2. Revisit is tasking-dependent, typically 1–4.5 days at mid-latitudes. Commercial licensing applies.
- PRISMA (ASI): Italian Space Agency hyperspectral mission: 239 contiguous bands from 400 nm to 2505 nm at 30 m spatial resolution, ~29-day revisit. The full spectral curve allows library matching against known material spectra rather than index ratios, which reduces confusion between weathered materials. Available through ESA Third Party Missions.
What a floating roof gives away
A metal roof and a concrete roof can share nearly identical red, green and blue reflectance values. Add a shortwave-infrared band centred near 2200 nm and they diverge sharply. Metal reflects strongly across SWIR; concrete absorbs more, particularly where clay minerals are present in the aggregate. Asphalt sits lower still. Clay tile has a characteristic absorption feature near 2200 nm tied to its aluminium hydroxyl bonds. These are not subtle differences visible only in laboratory conditions. They appear in satellite data collected from 786 km altitude.
The physical principle is straightforward. Surface materials have distinct molecular bond structures that absorb electromagnetic energy at predictable wavelengths. Visible bands (roughly 400–700 nm) respond mainly to colour pigments and surface roughness. SWIR bands (1000–2500 nm) respond to mineralogy, moisture content and organic compounds. A classification built on visible data alone is classifying colour. One built on SWIR is classifying chemistry.
Indices that do the separating
Several published spectral indices are designed specifically for built-surface discrimination. The Normalised Difference Built-up Index (NDBI) uses Sentinel-2 bands B11 (1610 nm) and B8A (865 nm) to separate built surfaces from vegetation and water. The Built-up Area Extraction Index (BAEI) incorporates the red band to reduce confusion with bare soil. For roofing material sub-classification, ratios of SWIR-2 to SWIR-1 reflectance are particularly useful: metal roofs typically show ratios above 1.0, concrete near 0.9, and clay tile below 0.85, though these thresholds shift with weathering and local material composition.
Plastic sheeting, common in informal settlements and disaster-affected areas, has a distinctive flat SWIR response with relatively high NIR reflectance. Thatch behaves spectrally more like dry vegetation than a built material, which is useful for informal-settlement characterisation but means it is easily confused with dry grass at 10–30 m resolution. These ambiguities are not reasons to distrust the method; they are reasons to document the confusion matrix honestly in every classification product.
The 10-metre floor and what lies below it
Sentinel-2's SWIR bands resolve at 20 m. A single pixel at that scale in a dense city contains contributions from a roof, a wall shadow, a narrow alley and possibly a tree canopy. The pixel's spectral signature is the area-weighted average of all of them, which corresponds to no real material on the ground. This is the mixed-pixel problem, and it is the dominant source of classification error in urban surface mapping.
The practical consequence is that Sentinel-2 material classification works well at neighbourhood or district scale, where the question is 'what proportion of roofs in this ward are metal versus concrete?' rather than 'what is this specific building roofed with?' For parcel-level answers, WorldView-3 SWIR at 3.7 m is the appropriate sensor. Even there, roof edges, gutters and HVAC equipment introduce sub-pixel mixing. No satellite sensor currently available resolves individual roof tiles on a standard residential building. Buyers who need that level of detail should combine satellite classification with a ground-truth survey, not replace one with the other.
Weathered concrete and bare soil: the persistent confusion
Freshly poured concrete and dry bare soil are separable in SWIR. Concrete aged five or more years is not always. Weathering leaches calcium carbonate to the surface, and wind deposits fine soil particles into surface pores. The spectral signature of a weathered flat concrete roof in a semi-arid city can sit within the natural variability of surrounding bare soil. Attempts to separate them using only two or three SWIR bands frequently misclassify one as the other.
PRISMA's hyperspectral data partially resolves this. With 239 bands, spectroscopic library matching can identify the specific absorption features of calcite versus kaolinite versus quartz-rich soil, even in weathered surfaces. The trade-off is 30 m spatial resolution and infrequent revisit. For most urban mapping projects, the practical approach is to use Sentinel-2 for area-wide classification, flag pixels with ambiguous SWIR ratios as 'uncertain', and resolve those specific areas with either WorldView-3 tasking or targeted field verification. Pretending the ambiguity does not exist produces a confident-looking map that is wrong in exactly the places it matters most.
From material map to policy number
A surface material classification is not useful by itself. Its value comes from the parameters it unlocks downstream. Runoff coefficients vary from roughly 0.1 for thatch or permeable paving to 0.95 for sealed metal or asphalt roofing. A material map combined with a digital surface model and local rainfall data produces spatially explicit runoff estimates that drainage engineers can use directly. Urban heat island modelling needs albedo and thermal emissivity by surface type, both of which correlate with material class. Informal-settlement upgrading programmes need to know how much of a target area is covered by plastic sheeting versus corrugated iron versus permanent roofing before they can estimate materials costs.
Satellize runs material classification workflows on Sentinel-2 and Landsat open archives, with WorldView-3 SWIR tasking added where parcel-level resolution is required under client licence. The Tonga crop-estimation programme demonstrated the team's approach to spectral index calibration in tropical conditions, where atmospheric moisture and dense canopy create similar challenges to urban mixed-pixel environments. Output formats are GIS-ready raster layers with per-class confidence scores, not just a single classified image.
Cloud, seasonality and the archive as a resource
Urban surface materials do not change quickly, which is both a challenge and an advantage. The challenge: a single cloudy image is useless, but cloud is not random. Tropical cities can have cloud cover exceeding 70 percent of acquisition days. The advantage: because materials are stable, analysts can composite multiple acquisitions across a dry season to produce a near-cloud-free surface reflectance image. Sentinel-2's five-day revisit makes this tractable. Landsat's 30 m resolution is a harder constraint, but the archive depth back to 1984 (Landsat 5 TM, which carried SWIR bands) allows change analysis over decades.
Seasonal variation matters. Wet-season images show higher surface moisture, which suppresses SWIR reflectance across all material types and shifts classification thresholds. A material map calibrated on dry-season imagery should not be applied directly to wet-season data without recalibration. This is documented behaviour, not a defect. Any workflow that ignores it will produce systematic errors that look like spatial patterns but are actually temporal artefacts.
Typical figures
| Spatial resolution (Sentinel-2 SWIR) | 20 m (SWIR bands B11, B12); 10 m (visible/NIR bands used in composite indices) |
| Spatial resolution (Landsat 8/9 OLI SWIR) | 30 m |
| Spatial resolution (WorldView-3 SWIR) | 3.7 m (SWIR); 1.24 m (multispectral) |
| Revisit (Sentinel-2, equator) | 5 days (two-satellite constellation) |
| Revisit (Landsat 8+9 combined) | ~8 days at equator |
| Key spectral bands for material separation | SWIR-1 (~1610 nm), SWIR-2 (~2190 nm), NIR (~865 nm), Red (~665 nm) |
| Minimum classifiable roof area (Sentinel-2) | ~400 m² (one 20 m pixel); reliable classification requires several contiguous pixels |
| Archive depth | Sentinel-2 from 2015; Landsat SWIR (TM/ETM+/OLI) from 1984 |
| Cloud sensitivity | Optical only; cloud cover blocks acquisition. Multi-date dry-season compositing required in humid tropics |
| Typical delivery format | GeoTIFF raster (material class + confidence), vector polygon summary by administrative unit, CSV of area statistics |
Analytics Satellize can run
| Roofing material map by class | SWIR ratio classification and NDBI-family indices applied to atmospherically corrected surface reflectance (Sen2Cor or LaSRC) | GeoTIFF raster with 6–8 material classes and per-pixel confidence score; polygon summary by ward or district |
| Runoff coefficient surface | Lookup table mapping classified material to published runoff coefficient ranges (e.g. ASCE design values), area-weighted by pixel | Raster layer for direct import into hydrological models; tabular summary by catchment |
| Albedo and emissivity surface by material | Material-class-to-albedo lookup from published spectral libraries, validated against Landsat-derived broadband albedo | Input raster for urban heat island or energy-balance models; CSV of mean albedo by neighbourhood |
| Informal-settlement roofing audit | Plastic sheeting and thatch detection using NIR/SWIR spectral ratios; confusion-flagged pixels reported separately | Polygon layer of informal roofing extent with area statistics; uncertainty layer identifying ambiguous pixels |
| Multi-year material change detection | Bi-temporal or time-series comparison of classified composites using Landsat or Sentinel-2 archive; change flagged where class shifts exceed confidence threshold | Change raster and summary table showing material transitions (e.g. thatch to metal) by period and area |
| Parcel-level material classification (WorldView-3) | SWIR band ratio classification at 3.7 m resolution with object-based image analysis to delineate individual roof segments | Vector polygon layer with material class per roof segment; suitable for building permit or tax-base workflows |
| Weathered-concrete / bare-soil disambiguation report | Spectral angle mapping against PRISMA or published spectral libraries; ambiguous pixels flagged for field verification prioritisation | Uncertainty map with ranked list of locations requiring ground-truth; reduces unnecessary field visits by concentrating effort |
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.