Multitemporal spectral change detection of heritage stone surface weathering
Multispectral time series from Sentinel-2 and WorldView-3 can track surface darkening, biological crust growth, salt efflorescence and spalling on exposed masonry, but atmospheric correction quality determines whether the signal is real or artefact.
Sensors
- WorldView-3 (Maxar): 1.24 m panchromatic, 1.65 m multispectral VNIR (8 bands, 400–1040 nm), 3.7 m SWIR (8 bands, 1195–2365 nm). The SWIR bands are essential for discriminating gypsum and calcite efflorescence from silicate stone; revisit 1–4.5 days depending on latitude and tasking geometry.
- Sentinel-2 MSI (ESA): 10 m VNIR (bands 2–4, 8), 20 m red-edge and SWIR (bands 5, 6, 7, 8A, 11, 12). Revisit 5 days at the equator with both satellites. Pixel size limits use to large monument complexes (Angkor, Persepolis, Stonehenge landscape); individual wall panels are unresolvable. Free archive from 2015.
- Landsat 8/9 OLI-TIRS (USGS/NASA): 30 m multispectral VNIR and SWIR, 100 m thermal. Revisit 8 days per satellite, 16 days per sensor alone. Most useful for establishing long baselines: the Landsat archive extends to 1972, giving a 50-year change record no other freely available system can match.
- Sentinel-2 MSI red-edge bands (ESA): Bands 5 (705 nm), 6 (740 nm) and 7 (783 nm) at 20 m resolution. These bands are specifically sensitive to chlorophyll absorption, making them the primary discriminator for biological crust (cyanobacteria, lichens, algae) against bare stone, a distinction that broadband VNIR alone cannot reliably make.
What a reflectance curve gives away about a weathering stone
Every weathering process leaves a distinct spectral fingerprint. Biological crusts dominated by cyanobacteria and algae absorb strongly in the red (around 670 nm) due to chlorophyll and show a sharp reflectance rise into the near-infrared: the same red-edge signal used in vegetation indices. Black fungi and lichenised crusts flatten the entire VNIR reflectance curve and reduce albedo measurably. Salt efflorescence, by contrast, raises reflectance broadly across VNIR wavelengths and shows diagnostic absorption features in the SWIR: gypsum (CaSO4·2H2O) has well-documented absorption near 1450, 1750 and 2200 nm, which WorldView-3's SWIR bands straddle directly. Spalling and granular disaggregation expose fresh mineral surfaces with higher reflectance and altered texture, detectable as a localised increase in VNIR return relative to surrounding weathered stone.
The practical consequence is that spectral change detection is not simply a matter of watching brightness go up or down. Each process has a directional signature in a specific part of the spectrum. A time series that tracks band ratios, such as a red-edge normalised difference index for biological crusts or a SWIR band ratio for sulphate salts, is far more informative than a single broadband change metric. Published work on Angkor Wat and Borobudur has demonstrated this approach using both hyperspectral ground data and multispectral satellite imagery to validate spectral assignments.
Why atmospheric correction is the dominant error source, not resolution
Stone surface reflectance changes caused by weathering are often subtle: a biological crust thickening by a few millimetres may shift SWIR reflectance by 2–5 percentage points. Atmospheric path radiance and aerosol scattering can introduce errors of comparable or greater magnitude if not removed consistently across acquisitions. This is the central technical problem for multitemporal heritage stone analysis, and it is more consequential than pixel size for most applications.
The main correction schemes applied in heritage contexts are MODTRAN-based radiative transfer models (used in ENVI FLAASH and similar tools), the Sen2Cor processor applied to Sentinel-2 Level-1C data to produce Level-2A surface reflectance, and empirical line correction using field spectrometry of stable reference surfaces within the scene. Sen2Cor is well-documented and freely available, but its performance degrades in arid environments with high and variable dust loading, which is precisely the condition at many desert heritage sites such as Persepolis or Petra. Dark Object Subtraction, still widely used, is the least reliable method for subtle surface change work: it systematically underestimates path radiance in SWIR bands and should not be the sole correction approach. Published guidance from the Copernicus programme recommends cross-calibrating acquisitions using pseudo-invariant features (PIF), stable bright surfaces such as concrete aprons or quarried limestone blocks within the site boundary, to reduce inter-date radiometric inconsistency to below 1–2% reflectance units.
An additional complication is bidirectional reflectance distribution function (BRDF) effects. Stone surfaces are not Lambertian: viewing geometry and solar angle differences between acquisitions introduce apparent reflectance changes that can mimic or mask real weathering signals. For Sentinel-2, the European Space Agency publishes BRDF correction coefficients, and their application is recommended before any multitemporal analysis.
Resolution arithmetic: which sensor fits which monument
Sentinel-2 at 10–20 m per pixel is only useful where the feature of interest covers at least several contiguous pixels, meaning the minimum mappable unit is roughly 400–1600 m². That constraint rules out individual wall panels, column drums or carved reliefs, but it is workable for large complexes. The Angkor Archaeological Park covers approximately 400 km²; Persepolis's terrace platform is roughly 450 m × 300 m, which gives around 675 Sentinel-2 pixels at 20 m resolution, enough to detect zone-level darkening trends but not to attribute them to specific architectural elements.
WorldView-3 changes the equation substantially. At 1.24 m VNIR resolution, individual stone courses are resolvable on most large masonry structures. The 3.7 m SWIR pixels are coarser but still allow salt efflorescence mapping at the level of wall bays. The cost is coverage: WorldView-3 is a commercial tasking asset, so archive depth at any given heritage site is typically sparse unless the site has been systematically tasked. Landsat's 30 m pixels are most valuable for long-baseline trend detection, not for fine spatial discrimination. A practical programme usually combines all three: Landsat for decadal trend, Sentinel-2 for annual monitoring, and WorldView-3 for targeted high-resolution campaigns at priority zones.
Change detection methods that hold up in practice
Image differencing and principal component analysis of multitemporal stacks are the most common approaches in published heritage remote sensing studies. Image differencing on atmospherically corrected SWIR bands has been used to map salt efflorescence progression at Angkor, where the laterite and sandstone substrates have sufficiently distinct SWIR signatures to separate stone from biological and saline weathering products. The limitation is that differencing conflates all sources of change, including vegetation encroachment, shadow shifts and correction artefacts, so post-classification filtering against a stable-pixel mask is necessary.
Spectral mixture analysis, where each pixel's reflectance is decomposed into fractional contributions from endmembers (fresh stone, biological crust, salt, shadow), offers a more physically grounded alternative. Endmember spectra can be sourced from field spectrometry or from published spectral libraries. The method produces fractional abundance maps at each date, and change in fractional abundance is a more interpretable metric than raw band difference. The honest limit is that endmember selection is subjective and sensitive to the atmospheric correction applied: two analysts using different correction schemes may produce meaningfully different abundance maps from the same raw imagery.
For operational monitoring programmes, a simpler but effective approach is to track a small set of targeted band ratios through time at fixed sample polygons drawn over known material zones. A red-edge chlorophyll index (such as the Sentinel-2 red-edge NDVI variant) tracks biological crust; a SWIR ratio near 1650/2200 nm tracks sulphate salt abundance. Plotting these through a multi-year Sentinel-2 time series produces a legible trend chart that conservation teams can act on without specialist remote sensing training.
Honest limits of the method
Cloud cover is an obvious constraint at tropical heritage sites. Angkor receives over 1400 mm of rain annually, concentrated in a monsoon season that can produce cloud-free Sentinel-2 acquisitions only intermittently between May and October. Compositing strategies (selecting the lowest-cloud-fraction image per quarter) help, but they introduce temporal averaging that blurs seasonal weathering dynamics.
Spatial resolution prevents the detection of early-stage micro-cracking, thin salt films below a few millimetres, or point-source biological colonisation on individual stones. Satellite spectral change detection is a screening tool: it identifies zones where ground inspection is warranted, not a substitute for close-range photogrammetry or petrographic analysis. The method also cannot distinguish reversible from irreversible change without ancillary data; a temporary algal bloom and a permanent lichen crust may produce similar short-term spectral shifts.
Satellize applies this class of analysis on Sentinel-2 and Landsat archives, with WorldView-3 tasking added on client licence, using the same atmospherically corrected time-series pipeline developed for the Kingdom of Tonga crop-estimation programme and adapted for heritage stone spectral indices.
Typical figures
| Spatial resolution (VNIR) | 10 m (Sentinel-2), 30 m (Landsat 8/9), 1.24 m (WorldView-3) |
| Spatial resolution (SWIR) | 20 m (Sentinel-2 bands 11–12), 30 m (Landsat OLI), 3.7 m (WorldView-3) |
| Revisit interval | 5 days (Sentinel-2 two-satellite constellation), 8 days per Landsat satellite, 1–4.5 days (WorldView-3, tasked) |
| Key spectral bands for weathering | Red-edge 705–783 nm (biological crust); SWIR 1650 nm and 2200 nm (sulphate salts, clay minerals) |
| Minimum mappable area (Sentinel-2) | Approximately 400–1600 m² (2–4 contiguous 20 m pixels) for reliable change detection |
| Atmospheric correction schemes | Sen2Cor (Sentinel-2 L2A), LaSRC (Landsat Collection 2 SR), MODTRAN-based FLAASH, empirical line with field spectrometry |
| Archive depth | Sentinel-2 from 2015; Landsat from 1972 (TM/ETM+/OLI); WorldView-3 from 2014 (site-dependent tasking history) |
| Radiometric accuracy target (post-correction) | Less than 2% reflectance units inter-date error, achievable with PIF normalisation on stable reference surfaces |
| Delivery formats | GeoTIFF surface reflectance stacks, band-ratio change maps (GeoTIFF/shapefile), time-series CSV per sample polygon, PDF conservation briefing |
Analytics Satellize can run
| Biological crust progression map | Multitemporal red-edge NDVI variant (Sentinel-2 bands 5 and 4) differenced across annual composites, masked against stable stone reference pixels | Annual GeoTIFF showing fractional change in biological crust index per 20 m cell, with zone-level trend summary in PDF |
| Salt efflorescence extent and intensity map | SWIR band ratio (WorldView-3 bands centred near 1650 and 2200 nm, or Landsat OLI bands 6 and 7) compared against gypsum spectral library endmembers | GeoTIFF abundance layer and shapefile of high-confidence efflorescence zones, updated per tasking cycle |
| Decadal surface darkening trend | Landsat Collection 2 surface reflectance time series (1984 to present) with LaSRC correction; linear trend fitted per pixel in VNIR broadband albedo | Trend magnitude raster (reflectance units per decade) and per-zone CSV time series for conservation reporting |
| Spalling and surface loss detection | Image differencing of atmospherically corrected VNIR composites combined with texture analysis (GLCM entropy) to identify newly exposed high-reflectance mineral surfaces | Change polygon shapefile with attributed change magnitude and date range, suitable for import into heritage GIS |
| Pseudo-invariant feature (PIF) correction report | Automated selection of stable bright reference pixels within scene; inter-date normalisation coefficients computed and applied across the full time stack before any change product is generated | Correction coefficient log and before/after reflectance comparison plots, provided as QA documentation alongside all analytic products |
| Priority zone alert | Threshold exceedance on monitored band-ratio time series at pre-defined sample polygons; alert triggered when quarterly change exceeds agreed conservation threshold | Email alert with map attachment and supporting spectral plot; threshold values agreed with client conservation team at programme outset |
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.