Identifying training sites?

I wont be able to give a great deal of information on this, but there are certain rules I think we need to follow for training sites:

• Training sites should be spread all over the image i.e. should be covering a large area and not just focussed in a region
• Just as Georeferencing, training sites should be identified at the corners and at points which give maximum sampling options
• The objective is to identify a set of pixels that accurately represents spectral variation present within each information region

The article below highlights some concepts regarding number of pixels per training site, and number of training sites per class for sampling (pg 2). Hawth's tools can be used to generate random stratified training points if you are using ArcGIS.

Improving image classification accuracy: a method to incorporate uncertainty in the selection of training sample sets