Hyperspectral Satellite Imagery and Linear Spectral Unmixing
Move beyond broad multispectral sensors with continuous 10 nm spectroscopy, spectral matching, and sub-pixel endmember unmixing.
Contiguous Narrow Bands Versus Broad Multispectral Sensors
Spaceborne hyperspectral imagers—including EnMAP, PRISMA, EMIT, and commercial fleets—capture 200 to 400+ contiguous bands across 400–2500 nm at ~10 nm sampling, replacing the wide wavelength gaps of multispectral sensors.
Continuous spectra expose sharp absorption troughs from clays, carbonates, iron oxides, hydrocarbons, and crop nutrients in the Shortwave Infrared (2100–2400 nm) that broad bands average out.
- VNIR (400–1000 nm): Pigments, chlorophyll, and iron oxide absorption.
- SWIR (1000–2500 nm): Clay minerals, carbonates, plastics, and cellulose.
- Narrow ~2300 nm SWIR features enable point-source methane plume detection.
Signal-to-Noise Ratio (SNR) and Water Vapor Removal
Splitting sunlight into 200+ narrow channels lowers photons per band, making SWIR Signal-to-Noise Ratio (> 150:1) critical. Atmospheric water-vapor bands near 1400 nm and 1900 nm absorb nearly all signal and must be masked before matching against USGS or ECOSTRESS spectral libraries.
- Mask 1350–1450 nm and 1800–1960 nm water-vapor bands before classification.
- Use Level-2A surface reflectance with verified smile/keystone calibration.
- Apply continuum removal to isolate diagnostic mineral absorption depths.
Spectral Angle Mapper (SAM) and Linear Spectral Unmixing
At 10–30 m pixel sizes, individual pixels blend soil, rock, and foliage. Spectral Angle Mapper (SAM) classifies pixels by vector angle regardless of terrain shading, while Linear Spectral Unmixing solves for exact fractional cover (0%–100%) of pure endmembers.
Fusing Hyperspectral Cubes with High-Res Optical Basemaps
Teams pair 10–30 m hyperspectral mineral or methane maps with 0.5 m optical scenes from Earth SAT Data to pinpoint exact drill pads, tailings boundaries, or pipeline infrastructure.
Multispectral vs Spaceborne Hyperspectral Sensor Characteristics
| Dimension | Multispectral (e.g., Sentinel-2) | Spaceborne Hyperspectral (EnMAP / EMIT) | Impact on Analytics |
|---|---|---|---|
| Spectral Bands | 4 – 13 discrete bands | 200 – 400+ contiguous bands | Resolves narrow chemical & mineral shapes |
| Bandwidth (FWHM) | 15 – 175 nm | 6 – 12 nm | Separates kaolinite vs alunite vs calcite |
| Spatial Resolution (GSD) | 0.3 m – 10 m | 10 m – 30 m | Sub-pixel unmixing resolves mixed cover |
| Data Volume per Scene | 50 MB – 800 MB | 1.5 GB – 6.0 GB | Requires PCA/MNF reduction or COG tiling |
| Primary Limitation | Blind between band gaps | SWIR SNR, water-vapor masks | Needs rigorous L2A atmospheric correction |
Commercial hyperspectral fleets increasingly target 5–15 m GSD alongside 30 m EnMAP and 60 m EMIT public archives.
Worked Illustration: Two-Endmember Linear Spectral Unmixing
A 30 m pixel measures Red (670 nm) = 0.144 and NIR (850 nm) = 0.340. Pure Green Canopy has (Red = 0.040, NIR = 0.500) and Bare Soil has (Red = 0.300, NIR = 0.100). We solve for vegetation fraction f_veg.
- Set up Red mixture equation: 0.144 = f_veg(0.040) + (1 - f_veg)(0.300).
- Solve for f_veg: 0.144 = 0.300 - 0.260 x f_veg -> f_veg = 0.156 / 0.260 = 0.60 (60%).
- Compute soil fraction: f_soil = 1 - 0.60 = 0.40 (40%).
- Verify NIR check: 0.60(0.500) + 0.40(0.100) = 0.340 (matches observed NIR).
Calculated Result & Takeaway: Linear unmixing decomposes the 30 m mixed pixel into 60% green vegetation and 40% bare soil with zero residual error.
Hyperspectral Workflow Verification Checklist
- Order L2A surface reflectance rather than L1B radiance for library matching.
- Drop water-vapor absorption windows (~1350–1450 nm and ~1800–1960 nm).
- Confirm SWIR SNR exceeds 150:1 for clay or hydrocarbon detection.
- Resample reference spectral libraries to the sensor's exact FWHM.
- Apply continuum removal prior to measuring SWIR absorption depth.
- Overlay fractional maps onto sub-meter optical imagery for validation.
Frequently Asked Questions
Why do hyperspectral satellites typically have 10–30 m GSD?
Dividing sunlight across 200+ narrow channels requires larger ground pixels to gather enough photons for acceptable SNR.
Can hyperspectral imagery distinguish specific mineral or plastic types?
Yes. Diagnostic SWIR absorption peaks between 1700 nm and 2400 nm separate clay minerals, carbonates, and polymers.
What is Continuum Removal?
Dividing a spectrum by its convex hull normalizes overall brightness so absorption band depths compare cleanly.
Can standard GIS tools open hyperspectral GeoTIFFs?
Yes. GDAL, Python rasterio, QGIS, and ArcGIS Pro read multi-band hyperspectral COGs efficiently via band subsets.
Primary References & Standards
- USGS Spectral Library Version 7 (Kokaly et al., 2017).
- DLR / GFZ EnMAP Hyperspectral Mission Level-2A Product Specification.
- NASA JPL EMIT L2A Surface Reflectance ATBD.