Minimum Noise Fraction (MNF)#
MNF reduces a hyperspectral cube to a smaller set of bands ordered by signal-to-noise ratio rather than by raw variance (as PCA does), so the first MNF bands hold the cleanest information and the last hold mostly noise.
How it works#
Estimate noise. Noise is approximated with the shift-difference method: each pixel is subtracted from its neighbor one row below. Adjacent pixels should be nearly identical, so what remains is largely sensor noise.
Whiten by noise. The noise covariance is computed and used to build a whitening transform that rescales every band so the estimated noise has equal (unit) variance in all directions.
Eigendecomposition. A standard eigendecomposition (PCA) is run on the whitened data covariance. Because the data is already noise-normalized, the resulting components are ranked by signal-to-noise ratio in descending order.
Project. The original cube is projected onto the top N eigenvectors, producing a new dataset with N MNF bands.
The result is added as a new dataset named MNF, Img: <source>.
Using the tool#
Pick a Dataset, set Num Components (the number of MNF bands to keep), and click OK. The spin box defaults to the maximum allowed for that dataset — the count of good bands, capped by the available noise samples — so that ceiling is visible at a glance.
Click View Past Results to reopen any previous run and view its scree plot (eigenvalue vs. component index), which helps you judge how many components are worth keeping.