I couldn’t find any technical description of what does unsupervised classification algorithm do with multiple source bands. I found that the classification results made by multiple source bands are indeed better than ones made with a single band, but I couldn’t find a clear justification of which bands to be chosen as sources.
(I’m working with Dual polarization images so I tried to combine the two polarizations in several ways e.g. average, ratio )
If you work with unsupervised classifiers, I’d say the more input layers the better.
Think of the n-dimensional feature space concept:
The clustering below is based on three axes (three input rasters). If it was only two dimensions blue/purple or red/yellow wouldn’t form such distinct clusters.
If some bands are not ‘useful’ for the classification (e.g. they show redundant information with another layer) their inpact on the result won’t be high.
I found nice slides about the topic:
www.utsa.edu/LRSG/Teaching/EES5083/L8-classif.ppt