Oil spill thickness and type estimation from optical sensors
SAR finds the spill; optical hyperspectral data tells you what kind of oil it is and, within limits, how thick. PRISMA and DESIS can separate crude, weathered oil, and water-in-oil emulsions that look identical in broadband imagery.
Sensors
- PRISMA (ASI): Italian Space Agency hyperspectral imager, 30 m spatial resolution, 400–2500 nm range across 237 contiguous bands at roughly 10 nm sampling. Revisit approximately 29 days at the equator for a single pass, though tasking can improve this. The spectral continuity is what separates it from multispectral alternatives: crude oil absorption features near 1200 nm and 1700 nm are resolvable.
- DESIS (DLR / Teledyne Brown Engineering): Hyperspectral sensor mounted on the ISS, 30 m resolution, 400–1000 nm in 235 bands at roughly 2.5 nm sampling. Visible and near-infrared only, so it cannot reach the shortwave-infrared oil absorption features, but it resolves fluorescence-related reflectance differences and pigment signatures in the 500–700 nm window that distinguish thin sheens from thicker oil.
- Sentinel-2 MSI: 13 multispectral bands, 10–60 m resolution depending on band, 5-day revisit at mid-latitudes with both satellites. Broad bands cannot resolve diagnostic oil absorption features directly, but band ratios (particularly B4/B3 and B8/B3) and sun-glint modelling can discriminate thick emulsions from thin sheens under favourable geometry. Freely available via Copernicus Data Space.
- Landsat 8 / 9 OLI: 30 m multispectral, 16-day revisit per satellite (8-day combined), coastal aerosol band at 443 nm useful for thin-sheen detection. Similar broadband limitations to Sentinel-2, but the archive back to 2013 (OLI) and the cross-calibrated pair allow multi-temporal analysis of chronic spill sites.
What spectral reflectance actually reveals about oil
Oil on water changes the surface's optical behaviour in two distinct ways. Thin sheens, typically less than 0.1 mm, act as interference films: they produce the iridescent colours visible to the naked eye and shift reflectance in the blue-green window in ways that depend on film thickness and refractive index. Thicker accumulations, from roughly 0.1 mm upward, absorb and scatter light differently from water, suppressing near-infrared reflectance and introducing absorption features tied to specific hydrocarbon bond structures. Water-in-oil emulsions, the brown 'chocolate mousse' that forms as a spill weathers, have a distinctly higher reflectance in the visible and a flatter spectral shape than fresh crude.
Hyperspectral sensors exploit these differences directly. PRISMA's coverage to 2500 nm captures C-H overtone absorption bands near 1200 nm and 1730 nm that are diagnostic of aliphatic hydrocarbons. Published work using airborne hyperspectral data (AVIRIS and HyMap) showed that spectral angle mapping against library spectra could separate fresh crude, emulsified oil, and plant-based sheens with classification accuracies above 80 percent under good illumination conditions. PRISMA brings similar spectral resolution to spaceborne scale, though at coarser spatial resolution than airborne sensors.
Sun-glint: noise source and accidental signal
Sun-glint is the specular reflection of sunlight off a near-flat water surface. It is the dominant source of confusion in optical spill analysis and, simultaneously, the reason thin sheens are visible at all. A thin oil film damps capillary waves, producing a smoother surface that reflects more sunlight toward the sensor at certain viewing geometries. This is why sheens appear bright against darker surrounding water in some images and dark against glinting water in others.
Modelling the glint is not optional. The Cox-Munk model, which relates wind speed to sea-surface slope statistics, is the standard starting point. Sensors with multiple view angles, or scenes acquired at off-nadir geometries, allow partial separation of glint from oil signal. Sentinel-2 and Landsat lack along-track stereo, so glint correction relies on the Cox-Munk approach combined with ancillary wind data, typically from ERA5 reanalysis or concurrent Sentinel-1 wind retrievals. Residual glint after correction remains the largest source of false positives in thin-sheen detection from multispectral data.
Thickness retrieval: what the numbers can and cannot tell you
Thickness retrieval from passive optical data is genuinely hard. The interference-film model works in principle for sheens below roughly 0.01 mm, where film thickness and reflectance colour are geometrically related. Above that, multiple scattering within the oil layer breaks the simple relationship. Published retrieval algorithms (notably those developed using AVIRIS data over the Deepwater Horizon spill in 2010) estimated thickness classes rather than continuous values, grouping observations into thin sheen, silver sheen, rainbow sheen, thick dark oil, and emulsion categories. Absolute thickness uncertainty within each class was large, often a factor of two or more.
Spaceborne sensors add further complications. At 30 m pixels, sub-pixel mixing is common at spill edges, where thin and thick oil coexist within a single resolution cell. Atmospheric correction over water is difficult, particularly in coastal scenes with aerosol loading. Cloud cover eliminates optical observation entirely, a problem SAR does not share. The honest position is that optical data provides oil type classification and relative thickness ranking with reasonable confidence, and absolute thickness in tonnes per square kilometre with poor confidence unless calibrated against in-situ samples or airborne data acquired concurrently.
Hyperspectral classification workflows
The standard processing chain for PRISMA or DESIS data over a suspected spill runs: atmospheric correction (typically ATCOR or 6SV adapted for water targets), sun-glint modelling and removal, spectral unmixing or spectral angle mapping against a reference library, and class-conditional probability mapping. The reference library matters enormously. Crude oils from different basins have different aromatic content and therefore different absorption feature positions. A library built on North Sea crude performs poorly on Venezuelan heavy crude. Operators who do not know the source of the spill must either use a broad library with reduced discrimination or run multiple hypotheses in parallel.
DESIS's restricted range to 1000 nm limits it to the visible and near-infrared, where fluorescence-related reflectance differences and the 700 nm chlorophyll edge (useful for separating biological films from oil) are the primary discriminants. It cannot reach the SWIR absorption features that make crude-versus-emulsion separation cleaner. For scenes where DESIS is the only hyperspectral option, classification accuracy for thick emulsions versus fresh crude drops noticeably compared to PRISMA-based results.
Operational limits buyers should plan around
Cloud cover is the most immediate constraint. The Mediterranean, Gulf of Mexico, and North Sea, three of the highest-risk spill regions, all experience multi-day cloudy periods that can delay optical observation entirely. A response workflow that relies solely on optical data will have gaps. The practical answer is fusion with SAR for detection and extent, reserving optical hyperspectral tasking for the characterisation step once the spill location is known.
PRISMA's 29-day single-pass revisit means opportunistic coverage of a fresh spill is unlikely without deliberate tasking through ASI. DESIS on the ISS follows a non-repeating ground track with an inclination of 51.6 degrees, which excludes latitudes above roughly 52 degrees north and south and makes revisit scheduling complex. Neither sensor was designed for rapid-response maritime surveillance. Sentinel-2's 5-day revisit and open data access make it the practical first look for multispectral characterisation, with hyperspectral tasking layered on for enforcement-grade classification where the geometry and cloud allow.
Satellize integrates PRISMA and Sentinel-2 optical analysis with concurrent SAR and wind data in its spill characterisation pipeline, applying the same spectral unmixing and glint-correction methods used in published post-Deepwater Horizon studies. The Tonga crop-estimation programme demonstrated the same atmospheric correction and spectral classification infrastructure in a different application context.
Typical figures
| Spatial resolution (hyperspectral) | 30 m (PRISMA, DESIS) |
| Spatial resolution (multispectral) | 10–30 m (Sentinel-2 MSI); 30 m (Landsat 8/9 OLI) |
| Spectral range | 400–2500 nm (PRISMA); 400–1000 nm (DESIS); 443–2200 nm (Sentinel-2); 435–2300 nm (Landsat OLI) |
| Spectral sampling (hyperspectral) | ~10 nm (PRISMA); ~2.5 nm (DESIS) |
| Revisit (single sensor) | ~29 days (PRISMA, equatorial); non-repeating ISS track (DESIS); 5 days (Sentinel-2 pair); 8 days (Landsat 8+9 combined) |
| Minimum detectable sheen thickness (optical) | Interference films detectable from ~0.001 mm; thickness retrieval reliable only to broad class (sheen / rainbow / thick / emulsion) |
| Cloud limitation | 100% data loss under cloud cover; no passive optical penetration of cloud |
| Archive depth | PRISMA from 2019; Sentinel-2 from 2015; Landsat OLI from 2013 |
| Latency (tasked acquisition to delivery) | Typically 24–72 hours for tasked hyperspectral; Sentinel-2 open data available within ~3 hours of acquisition |
| Delivery formats | GeoTIFF classification maps, spectral class polygons (GeoJSON/Shapefile), PDF enforcement report |
Analytics Satellize can run
| Oil type classification map | Spectral angle mapping and spectral unmixing against hydrocarbon reference library, applied to atmospherically corrected PRISMA or DESIS imagery | GeoTIFF and GeoJSON polygon layer with per-pixel class labels (fresh crude, weathered oil, water-in-oil emulsion, thin sheen, biological film) |
| Relative thickness ranking | Interference-film model for sheens below 0.01 mm; reflectance-based class thresholding for thicker accumulations, following methods published in post-Deepwater Horizon AVIRIS literature | Classified raster with five thickness categories and associated uncertainty flags; not a continuous thickness map |
| Sun-glint corrected scene | Cox-Munk wave-slope model combined with ERA5 wind speed input and near-infrared deep-water reference pixels | Atmospherically and glint-corrected surface reflectance cube (GeoTIFF), suitable for downstream classification or independent analysis |
| Spill extent and type fusion product | Co-registration of Sentinel-1 SAR extent mask with optical classification; SAR provides cloud-independent boundary, optical provides type attribution within the boundary | Merged GIS layer with SAR-derived outer boundary and optical-derived type zones; PDF enforcement summary with acquisition metadata |
| Temporal change report for chronic spill sites | Multi-date Sentinel-2 band-ratio analysis (B4/B3, B8/B3) with consistent glint correction across archive scenes | Time-series chart of spill area and spectral class proportions; GeoTIFF stack; suitable for regulatory reporting or legal evidence packages |
| Spill source vessel correlation | Spatial and temporal back-projection of spill drift using Copernicus Marine Service ocean current data, overlaid against AIS vessel position records | Candidate vessel shortlist with position-time evidence; map of drift trajectory; flagged for human review, not automated attribution |
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.