SAM doesn't extract endmembers, but since its algorithm takes a single spectrum as input, it calls the input an 'endmember'.  Whether this input is actually an endmember depends on how they were defined. An Endmember is considered a pure material, typically taken in a lab.  More often, users will extract endmembers from an image, so they are 'mixing endmembers', not totally pure but the pixel is considered close to 100% of the material of interest.  They are often used in sub-pixel analysis where a user wants to find out the percent of material in each pixel.  There is a whole process that can be performed to find these endmembers since they can be difficult to find visually.  The extraction process will usually result in a single pixel spectrum of the spectrally pure materials in the scene.  Sometimes they can be small groups of pixels but mostly they are single spectra. Training data are most often groups of pixels which represent good examples of a material (or class).  They can also be single pixel spectra, but more often are groups of pixels.  They are used in whole-pixel analysis and classification.  Many classification algorithms, such as max likelihood, require several pixels in the training class so covariance can be calculated.  So a single input spectrum (or endmember) would not be appropriate. In your case,  your input to SAM is a group of pixels in an ROI.  The average spectrum in this ROI is calculated and used in the angle comparisons.  If this average is a 'pure' representation of the class, then perhaps it would be considered an endmember.  If there is significant variance in the individual pixel spectra, its probably more of a training class.