Multispectral vs Hyperspectral: Bands, Cost and Free Data
Published Updated
A multispectral image records the light reflected from the ground in a handful of bands, typically 4 to 13, each a wide slice of the spectrum: blue, green, red, near infrared, a few in the shortwave infrared. A hyperspectral image records the same light in hundreds of narrow, contiguous bands, so every pixel carries a full spectrum instead of a few samples of it.
The difference in one line
Multispectral tells you that a field is green and stressed. Hyperspectral can tell you which crop, which mineral, which pigment.
| Multispectral | Hyperspectral | |
|---|---|---|
| Bands | 4 to 13 | 100 to 300 |
| Free satellites | Sentinel-2, Landsat | EnMAP, PRISMA, on request |
| Drone sensor | A few thousand euros | Tens of thousands |
| Data per scene | Hundreds of megabytes | Gigabytes |
| Answers | Vegetation, water, built, stressed | Which species, which mineral, which plastic |
Spectral resolution: what the extra bands buy
A hyperspectral image is a data cube: two spatial axes and one spectral axis. Every pixel holds a full spectrum, and the spectrum works like a fingerprint, because each material reflects light differently. A multispectral pixel holds a few samples of that spectrum, with gaps between them.
Spectral resolution is the width of each band and the spacing between bands. Industrial hyperspectral cameras space their bands 1 to 4 nm apart. Satellite sensors trade that for coverage: EnMAP samples every 6.5 nm in the visible and near infrared and every 10 nm in the shortwave infrared, which is fine enough to resolve mineral features around 2,000 nm.
More bands cost something. Hyperspectral cameras are among the most expensive cameras on the market, they scan slower, and a typical push-broom model images only about 1,000 pixels across the line. A multispectral camera is cheaper and faster, which is why it wins when a few well-chosen bands are enough.
What each one is good for
Multispectral does almost everything a GIS team does with satellite imagery, for free. Sentinel-2 delivers a new 13-band scene of every point on land every five days, and its near-infrared band is enough to map crops, forest, water, burned areas and built-up extent. The tool is index arithmetic, NDVI first, then a classification in the Semi-Automatic Classification Plugin. The tree canopy post runs that route.
Hyperspectral adds discrimination inside a class. Two crops that look the same in every Sentinel-2 band have different spectra at 10 nanometre resolution. Mineral mapping, its original use, reads absorption features no multispectral band can see. The price is expensive sensors, heavy data, and free satellites that image on request at 30 m with a revisit measured in weeks.
Named sensors and their bands
| Sensor | Type | Bands | Pixel size | Revisit |
|---|---|---|---|---|
| Sentinel-2 | Multispectral | 13 | 10, 20 and 60 m | 5 days at the equator |
| Landsat 8 | Multispectral | 9 (OLI) plus 2 thermal (TIRS) | 30 m, 100 m thermal | 16 days |
| EnMAP | Hyperspectral | 228, 420 to 2,450 nm | 30 m | 27 days at nadir, 4 days off-nadir |
| PRISMA | Hyperspectral | 239, 400 to 2,505 nm | 30 m | On request |
| EMIT | Hyperspectral | 285, 381 to 2,493 nm | 60 m | Dust-source regions, 52° N to 52° S |
The Sentinel-2 swath is 290 km wide. EnMAP and PRISMA both image a 30 km swath, so one scene covers a district, not a country. EnMAP data is free and open.
Which one for which job in QGIS
- Land cover, crop type, forest extent, flood extent. Multispectral. Free, frequent, and the QGIS classification tools were built for it.
- Crop stress before it shows, species, minerals, contamination, plastics. Hyperspectral, from a drone for a site, from EnMAP or PRISMA for a region.
- Anything that needs to see a small object. Neither. Both are pixel sciences at 10 to 30 m. A roof, a tree crown or a pool is a shape, read from high-resolution imagery in three bands. AI Segmentation, the plugin we make at TerraLab for QGIS, works that way: it reads the RGB orthophoto and returns one polygon per object.
, works that way: it reads the RGB orthophoto and returns one polygon per object.Questions people ask
Is Sentinel-2 multispectral or hyperspectral?
Multispectral: 13 bands at 10, 20 and 60 m. The free hyperspectral missions are EnMAP, PRISMA and NASA's EMIT.
What is the red edge?
The steep rise in a leaf's reflectance between about 680 and 750 nanometres. Its position shifts with plant stress, which is why Sentinel-2 carries three narrow bands across it.
Does QGIS support hyperspectral data?
Yes. A hyperspectral scene is a raster with a few hundred bands and every raster tool reads it. EnMAP-Box, a QGIS plugin from the EnMAP team, is built for that case.
How much does a multispectral camera cost?
A drone multispectral sensor costs a few thousand euros, a hyperspectral one tens of thousands. Satellite multispectral data from Sentinel-2 and Landsat costs nothing.
What does a hyperspectral image look like?
Like a normal photo, until you click a pixel. Then you get a curve of 200 or more values instead of three. An EnMAP scene, for example, is a 30 km wide strip with 228 bands per pixel. Outside the Earth sciences, the same technique sorts waste, inspects food and checks pharmaceutical production.
The QGIS AI hub collects what AI can and cannot do inside QGIS. The imagery vocabulary continues on the orthomosaic vs orthophoto page.
