SAM doesn't require that you find endmembers via the 'hourglass' approach but it is a good method to use, particularly for hyperspectral datasets which have a lot of correlated bands. The purpose of using the hourglass approach (MNF->Pixel purity->n-D visualizer) is to help you reduce the redundant (correlated) information in your bands so that you may find the purest examples of material (endmembers) in your image to use in classification. It's particularly useful if you wish to unmix pixels.
The result of running MNF, then PP are pixels of the most extreme (and potentially 'pure' ) materials in your image. You load these pixels into the n-D visualizer in order to view them in their n-D space to find the 'outliers' which are most likely your true endmembers of the image. I am sure the results seems strange until you become familiar with the tools but perhaps you can explain what 'strange' means?
If you are running these steps in the Hourglass Wizard, then you have the option to let the tool extract the most extreme pixels based on statistics. If you select this option, you can use the 'export to spectral library' which can then be used in the SAM classifier. If you manually select pixels yourself in the n-D visualizer, you can export the classes to the ROI tool to use as classification ROIs, or plot the mean spectrum from the ROI and save the plot as a spectral library to use in SAM. Does this help?
|