Spectral Analytics
Spectral Analytics treats each pixel as a spectrum, a curve of reflectance across wavelengths, rather than as a handful of independent band values. That shift matters when your imagery has many bands, and it is the natural way to work with hyperspectral data.
Two questions dominate this sub toolbox. What material is this pixel, judged by the shape of its spectrum? And when a pixel contains several materials at once, in what proportions?
Start here
Section titled “Start here”Linear Spectral Unmixing
Section titled “Linear Spectral Unmixing”The answer to the mixed pixel problem. A 20 metre pixel over a suburb is not a house or a garden, it is a mixture of both, and no classifier that must choose one class will describe it correctly.
Unmixing assumes each pixel’s spectrum is a combination of a few pure material spectra, the endmembers, and solves for their proportions. The result is one raster per endmember, holding the fraction of the pixel that material occupies.
The endmembers are the whole difficulty. They can come from a spectral library, or from the purest pixels of your own image, which is often the more reliable route since they share the same acquisition conditions.
Spectral Angle Mapper
Section titled “Spectral Angle Mapper”Classifies pixels by comparing the shape of their spectrum with reference spectra, measuring the angle between them as vectors.
Its virtue is that the angle ignores overall brightness. The same material in sunlight and in shadow has the same spectral shape and different brightness, so where a distance based classifier sees two classes, the spectral angle sees one. That is why it is the standard method for mineral mapping and for any scene with strong illumination contrast.
Spectral Library Matching does the same with several similarity measures rather than the angle alone.
Continuum Removal
Section titled “Continuum Removal”A normalisation step used before spectral matching. It fits the upper envelope of the spectrum, the continuum, and divides the spectrum by it, which leaves the absorption features standing out clearly.
Those features are the diagnostic part: a mineral, a plant stress, a water content shows up as a dip at a specific wavelength, and continuum removal is what makes that dip comparable between spectra of different overall brightness.
Minimum Noise Fraction
Section titled “Minimum Noise Fraction”The hyperspectral equivalent of a principal component analysis, and better suited to it. It estimates the noise first, whitens it, then transforms, so the resulting components are ordered by signal to noise ratio rather than by variance alone.
The practical use is dimensionality reduction. A 200 band hyperspectral cube usually carries its information in the first ten or twenty MNF components, and working on those is faster and cleaner than working on the raw bands.
All the tools
Section titled “All the tools”| Tool | What it does |
|---|---|
| Linear Spectral Unmixing | Decomposes each pixel into the proportions of a few endmember materials. |
| Spectral Angle Mapper | Classifies pixels by the angle between their spectrum and reference spectra. |
| Spectral Library Matching | Matches pixel spectra to a reference library using several similarity metrics. |
| Continuum Removal | Normalises spectra by their upper envelope, isolating absorption features. |
| Minimum Noise Fraction | Transforms the bands into components ordered by signal to noise ratio. |
These tools reward many bands. With four bands the spectral shape is barely a shape. They come into their own on hyperspectral data, and remain useful on multispectral imagery for unmixing.
Unmixing fractions should sum to one. Checking that is the quickest sanity check of an unmixing result. Fractions that consistently miss the total usually mean an endmember is missing from the set.