You cannot use Maximum likelihood, mahalanobis distance or similar methods that require the calculation of covariance and matrix inversion. This is because a spectrum,whether collected from a single pixel or calculated from a group, is a single mean value. To calculate covariance requires multiple samples per training class so it is only designed to work with ROIs which typically comprise spectra from multiple pixels. I would try using a method such as SAM which does not require the calculation of covariance. If your images have spatial reference, you should be able to restore ROIs on each image - assuming they cover the same spatial extents.

Thanks Mari. You are very helpful, I really appreciate your insight. I think that restoring ROI could work between datasets I have from the same scene. Much thanks, again! Alex