Hi,
Could you please elaborate on what you mean by the SD values being skewed for 16-bit data? In general I would expect 1 SD in 16-bit imagery to be a larger value than 1 SD in 8-bit imagery. Signatures derived from those images would be similarly affected in range. For example, here's the stats for some signatures I quickly extracted in a 16-bit (actually 11-bit) IKONOS MSI image:
..which seem about correct. The Built-up SDs are quite large because I included quite a high degree of variation within the sample I took.
Which is a good point to remember - when looking at 16-bit imagery you probably aren't seeing all the variation that is present in the area you are looking at because your display is using 8-bit color guns (three 8-bit color guns and an overlay channel are usually thought of as a colordepth of 32-bit). So you may want to use the display extent to stretch the display of the data when collecting signatures so that you can see more of the variation you might be including. Displaying the image using the Image Chain mode and turning on Discrete Update is a good, efficient way of doing this so the stretch is adapted to the displayed image range as you roam around collecting signatures.
How did you generate the signatures? Did you use the Region Grow tool (which should result in normally distributed, unimodal signatures), or did you manually digitise a polygon (which might have inadvertently included outlier pixels which then resulted in too large a statistical range for the signature)? Or did you perhaps use existing polygon features? Or some other method?
Cheers