Spectral mixture analysis of surface material composition at heritage sites
Linear spectral unmixing applied to hyperspectral and SWIR imagery maps the areal proportions of construction materials across heritage site surfaces, supporting condition assessment without a single field visit.
Sensors
- ASI PRISMA: Hyperspectral imager covering 400–2500 nm in 239 contiguous bands at 30 m ground sampling distance, with a 30 km swath. Revisit is roughly 29 days at the equator, shorter at higher latitudes. The full VNIR–SWIR range is sufficient to resolve the diagnostic absorption features of calcite, fired clay and gypsum plaster.
- WorldView-3: Eight SWIR bands (1195–2365 nm) at 3.7 m resolution, plus eight VNIR multispectral bands at 1.24 m. The SWIR bands directly straddle the hydroxyl and carbonate absorptions that separate mudbrick from limestone from fired brick. Tasked revisit can be less than one day at mid-latitudes.
- DESIS (ISS): Hyperspectral sensor on the International Space Station covering 400–1000 nm at roughly 2.55 nm spectral sampling and 30 m spatial resolution. VNIR-only, so it cannot resolve SWIR carbonate features, but it adds useful discrimination of iron-oxide weathering crusts and organic soil staining.
- Sentinel-2 MSI: Three SWIR bands at 20 m resolution (1610 nm and 2190 nm most useful) with a five-day revisit at mid-latitudes. Spectral resolution is coarse compared with PRISMA or WorldView-3 SWIR, so unmixing endmembers must be broad material classes rather than specific mineral phases. Useful for site-wide mapping and change monitoring between targeted acquisitions.
What spectral unmixing actually measures
Every pixel in a satellite image is a mixture. A 30-metre PRISMA pixel covering a ruined Byzantine basilica contains fragments of limestone ashlar, fired roof tile, gypsum plaster, mudbrick and windblown soil, all averaged into one spectral vector. Linear spectral unmixing treats that vector as a weighted sum of pure material spectra (endmembers) and solves for the fractional abundance of each. The physics is straightforward: if the mixing is at a scale finer than the pixel and the surfaces are optically flat, the model is linear and the solution is constrained to fractions that sum to one.
The practical question is whether the endmembers are spectrally separable. For common heritage construction materials, the answer is yes, provided you have SWIR coverage. Calcite-based limestone and marble show a strong absorption near 2340 nm. Fired clay brick has a broad hydroxyl feature near 2200 nm and a characteristic iron-oxide slope in the visible. Mudbrick (unfired clay-rich sediment) has a shallower hydroxyl feature at a slightly different position. Gypsum plaster has a distinctive doublet near 1750 nm. Sentinel-2 can hint at these differences; PRISMA resolves them; WorldView-3 SWIR places them at 3.7 m with enough spatial detail to follow individual wall lines.
Where published studies put the method to work
The peer-reviewed record is specific enough to set realistic expectations. Studies at Roman-period sites in North Africa and the Levant have used PRISMA and its predecessor CHRIS/PROBA to map the spatial extent of opus signinum flooring, fired brick hypocaust debris and limestone rubble within the same ruin field, achieving classification accuracies in the 80–90 per cent range against field-verified reference points, though those figures depend heavily on endmember quality and site-specific atmospheric correction. At Islamic-period mudbrick sites in Iran and Iraq, researchers have used WorldView-3 SWIR to distinguish standing mudbrick from collapsed mudbrick from redeposited alluvial soil, a distinction that is nearly invisible in panchromatic or standard RGB imagery.
Byzantine sites in the eastern Mediterranean present a particular challenge because gypsum plaster, lime plaster and weathered limestone can have similar broadband reflectance while differing in narrow SWIR features. This is exactly the regime where PRISMA's 239-band coverage earns its keep over a four-band multispectral sensor. Published work from the PRISMA mission science team documents detection of gypsum efflorescence on standing masonry, which matters for conservation planning because gypsum crystallisation is a primary driver of stone spalling.
Honest limits: resolution floors, cloud and the endmember problem
PRISMA's 30 m pixel is the binding constraint at most standing-monument sites. A Roman forum column is roughly 1.5 m in diameter; it will never fill a PRISMA pixel. The method works at the site scale, mapping broad material zones across a ruin field, not at the architectural-element scale. WorldView-3 SWIR at 3.7 m gets closer to individual wall courses but is still not a substitute for close-range hyperspectral scanning of a single ashlar face.
Cloud cover is a genuine problem at Mediterranean coastal sites in winter and at highland sites year-round. PRISMA has no cloud-penetration capability. A single clear acquisition may be all that is available for a given season, and atmospheric haze degrades SWIR retrieval even in nominally cloud-free conditions. Careful atmospheric correction using radiative transfer models (ATCOR, 6S or similar) is not optional.
The endmember problem is underappreciated. Spectral libraries built from laboratory samples do not always match field spectra because weathering, biological crusts, dust coatings and moisture all shift reflectance. Best practice is to extract endmembers from the image itself using algorithms such as N-FINDR or SMACC, then validate a subset against field or drone-based reference measurements. Without that validation step, abundance maps can look plausible while being materially wrong.
From abundance fractions to a conservation decision
A fractional abundance map is not itself a conservation output. The translation step requires domain knowledge. A high mudbrick fraction in a zone previously recorded as stone construction may indicate collapse and sediment accumulation, or it may indicate that the original survey was wrong. A rising gypsum fraction on a previously stable limestone facade is a more unambiguous signal: salt crystallisation is progressing and intervention timelines should be reviewed.
Multitemporal stacking adds the most value. If WorldView-3 SWIR acquisitions are taken annually or after significant rainfall events, the change in fractional abundance between epochs can be mapped. A 10 per cent increase in soil fraction across a previously consolidated mudbrick platform is a quantitative proxy for surface erosion, comparable to what a field inspector would record qualitatively as 'significant deterioration'. That number can feed directly into a condition-scoring matrix of the kind used by ICOMOS and national heritage agencies.
Satellize applies this workflow, including atmospheric correction, image-derived endmember extraction and multitemporal differencing, as part of its heritage analytics service.
What the data cannot replace
Spectral unmixing from orbit maps surface composition. It says nothing about stratigraphy, structural integrity or subsurface material. A pixel showing 60 per cent fired brick abundance could be a well-preserved hypocaust floor or a thin scatter of tile fragments over bare earth. Ground-truth, even a single targeted field visit, resolves that ambiguity and dramatically improves the confidence of the abundance estimates across the rest of the site.
Pigment identification on painted surfaces is beyond the reach of 30 m hyperspectral data and is handled separately by close-range methods. The sibling page on hyperspectral pigment and weathering mapping at open-air rock-art sites covers that narrower problem. What satellite spectral unmixing does well is the site-wide material inventory at a scale and frequency that no ground survey programme can match economically.
Typical figures
| Spatial resolution (hyperspectral) | 30 m (PRISMA, DESIS); 3.7 m SWIR / 1.24 m VNIR (WorldView-3) |
| Spectral range | 400–2500 nm (PRISMA); 400–1000 nm (DESIS); 1195–2365 nm SWIR + 427–918 nm VNIR (WorldView-3) |
| Number of spectral bands | 239 contiguous (PRISMA); 8 SWIR + 8 VNIR (WorldView-3); 235 VNIR (DESIS) |
| Revisit (hyperspectral) | ~29 days (PRISMA, untasked); <1 day possible (WorldView-3, tasked); variable (DESIS, ISS orbit) |
| Minimum mappable material zone | ~1 ha at 30 m resolution (PRISMA); ~15 m² at 3.7 m resolution (WorldView-3 SWIR) |
| Key diagnostic absorptions | Calcite ~2340 nm; fired clay hydroxyl ~2200 nm; gypsum doublet ~1750 nm; iron oxide visible slope 450–700 nm |
| Cloud sensitivity | Full; no cloud penetration. Atmospheric correction required even in cloud-free conditions |
| Archive depth | PRISMA: 2019–present; WorldView-3: 2014–present; Sentinel-2: 2015–present |
| Typical delivery format | GeoTIFF fractional abundance layers per endmember; vector polygons of dominant material zones; CSV change statistics per site polygon |
| Atmospheric correction requirement | Mandatory; radiative transfer models (e.g. ATCOR, 6S) recommended; surface reflectance product preferred over top-of-atmosphere |
Analytics Satellize can run
| Site-wide material abundance map | Linear spectral unmixing with image-derived endmembers (N-FINDR or SMACC); constrained least-squares solution | GeoTIFF stack, one band per endmember (limestone, fired brick, mudbrick, plaster, soil), fractional values 0–1 |
| Endmember spectral library | Automated pure-pixel extraction from PRISMA or WorldView-3 SWIR imagery, validated against published USGS spectral library references | CSV spectral library with wavelength positions, reflectance values and material labels; reusable across subsequent acquisitions |
| Multitemporal material change map | Pixel-wise differencing of fractional abundance layers between two or more epochs; change thresholded at analyst-validated significance level | GIS polygon layer of zones with statistically significant material-fraction change; PDF report with before/after abundance histograms |
| Gypsum efflorescence extent mapping | Targeted unmixing using gypsum doublet feature at 1750 nm; PRISMA band depth index combined with abundance fraction | Vector polygon layer of gypsum-positive surface zones with fractional abundance values; suitable for direct input to ICOMOS condition-scoring matrices |
| Mudbrick versus alluvial soil discrimination | Hydroxyl feature position analysis (2200 nm vs 2160 nm shift) combined with iron-oxide visible slope ratio; WorldView-3 SWIR band ratios | Classified raster distinguishing standing/collapsed mudbrick from redeposited alluvial sediment; confidence layer based on spectral angle distance |
| Condition-scoring input table | Zonal statistics of fractional abundance per site management polygon; comparison against baseline epoch | CSV table of material fractions and change rates per zone, formatted for direct import into heritage management databases |
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.