What is hyperspectral imagery?

Hyperspectral imagery is imagery that records a continuous spectrum for every pixel, typically 100 to 300 narrow bands 5 to 10 nanometers wide across visible, near-infrared, and shortwave infrared light. Multispectral imagery records 3 to 15 broad bands. The extra spectral detail lets analysts identify minerals, crops, plastics, and gases by their reflectance signature.

What hyperspectral imagery is

Hyperspectral imagery comes from an imaging spectrometer, a sensor that splits incoming light into hundreds of narrow, contiguous wavelength bins and records each one for every pixel.

The result is a data cube. Two dimensions are the image; the third is wavelength, and a single pixel is a full reflectance curve from about 400 to 2,500 nanometers.

The defining property is spectral resolution. Bands are narrow enough and close enough together to resolve absorption features a few nanometers wide, which is what makes a material identifiable from its spectrum alone.

Multispectral imagery, by contrast, samples a handful of broad bands chosen for general purposes. Sentinel-2 has 13; Landsat has 11. Those bands separate vegetation from soil and water but cannot separate one mineral from another.

Orbital hyperspectral sensors include NASA’s EMIT on the space station, DLR’s EnMAP, Italy’s PRISMA, and Planet’s Tanager. Airborne sensors such as AVIRIS-NG fly at finer resolution over smaller areas.

What hyperspectral imagery does

Hyperspectral imagery identifies what a surface is made of, where multispectral imagery can only say what class it belongs to.

Mining companies map alteration minerals across a prospect from their absorption features, narrowing where to drill.

Methane monitoring programs detect plumes from the gas’s absorption near 1,650 and 2,300 nanometers, then quantify emission rates with wind data.

Agronomists separate crop varieties and detect disease and nutrient stress weeks before it is visible in color imagery.

Environmental agencies classify plastics on beaches and in rivers, and map invasive plant species by their spectra.

Water quality teams estimate chlorophyll, suspended sediment, and harmful algal blooms from the shape of the water’s reflectance curve.

Each use trades coverage for detail. Hyperspectral scenes are narrower and rarer than multispectral ones, so most workflows use hyperspectral imagery to calibrate or target what multispectral imagery then monitors.

How hyperspectral imagery works

A hyperspectral scene is stored as a multi-band raster with band count in the hundreds, usually as GeoTIFF, NetCDF, or ENVI format with a header listing each band’s center wavelength.

Sensor	Bands	Spectral range	Pixel size
NASA EMIT	285	380 to 2,500 nm	60 m
DLR EnMAP	224	420 to 2,450 nm	30 m
ASI PRISMA	239	400 to 2,505 nm	30 m
Planet Tanager	400+	380 to 2,500 nm	30 m
Sentinel-2 (multispectral)	13	443 to 2,190 nm	10 to 60 m

Processing starts with radiometric calibration and atmospheric correction, which are harder than for multispectral imagery because water vapor and CO2 absorb strongly in specific narrow bands.

Analysis then takes one of three routes. Spectral matching compares each pixel’s curve to a library of known material spectra, and band ratios isolate a single absorption feature. Machine learning models consume the full cube.

Dimensionality is the practical problem. Adjacent bands are highly correlated, so principal component analysis or band selection often reduces 200 bands to 10 or 20 before classification.

File size follows. A 60 m EMIT scene of 285 bands runs to about 1.5 GB, and a 30 m sensor over the same area produces four times as much.

Output is a classified raster, an abundance map per material, or a vector layer of detected features such as plumes or outcrops.

Hyperspectral imagery in Wherobots

WherobotsDB reads hyperspectral imagery as a multi-band raster and runs band selection and band math in SQL, in parallel across every scene in a table.

RS_NumBands reports how many bands a scene holds, RS_Band extracts a subset, and RS_MapAlgebra applies an expression across any bands by index. A methane band ratio or a narrow-band vegetation index is one statement.

Scenes load as out-db rasters with RS_FromPath, so a 1.5 GB cube in S3 is read only in the bands and tiles a query touches. Nothing is downloaded first.

Havasu, the Apache Iceberg table format extended with a native raster type, holds hundreds of scenes as rows, with metadata columns for sensor, date, and wavelength list alongside the raster.

RasterFlow runs PyTorch models over multi-band mosaics through its bring-your-own-rasters path. The inference config takes a list of band indices as features, so a model trained on 20 selected bands reads only those.

Because raster and vector data share one engine, a detected plume or mineral outcrop joins to well pads, leases, or parcels in the same query.

The same functions handle multispectral imagery, so a workflow that targets with hyperspectral imagery and monitors with Sentinel-2 runs in one system.

-- Normalized difference between two narrow bands (by index) across all scenes
-- e.g. a red-edge chlorophyll index on an EnMAP cube
SELECT scene_id,
       RS_MapAlgebra(rast, 'D',
         'out = (rast[63] - rast[41]) / (rast[63] + rast[41]);', -9999.0) AS index_rast
FROM hsi.enmap_scenes
WHERE RS_NumBands(rast) = 224;

Related terms

• NDVI: the two-band index most multispectral work starts with
• Methane detection satellites: the main commercial use of orbital hyperspectral sensors
• Sentinel-2: the most used multispectral mission
• Remote sensing: the field both imagery types belong to

Read more from Wherobots

• RasterFlow models
• RasterFlow data models
• Bringing your own PyTorch model to RasterFlow
• RasterFlow is now available in Public Preview

Run band math on hyperspectral imagery in SQL with WherobotsDB at cloud.wherobots.com.