Your process is fairly complicated but it sounds like you have performed this kind of analysis before. The error you are running into with the ROI separability is essentially the same as described in this help article:
http://www.exelisvis.com/Support/Help...
There is something in your ROI data that is causing a singularity in the covariance matrix which stops the process. It could be duplicate bands or ROIs with bands that are too similar. You might try testing the classification process with just the original spectral data, then start adding in the products you created (you mentioned vegetation indices). I also think classifications on images with combinations of bands that represent the original reflectance data and products can cause strange behavior with some algorithms because the data ranges can be very different in those kinds of images. What classification algorithm are you using? Does it work with other such as SAM (typically it is the algorithms such as maximum likelihood that are sensitive to data ranges)?