I would use MTMF and only the endmembers/ROIs/classes I am interested in finding. Unlike linear spectral unmixing, MTMF does not require you include all possible image endmembers to find relative abundances within a pixel. MTMF is a special case of Matched filter in that it also calculates infeasibility to help assess if you have false positives:
http://www.harrisgeospati...atchedFiltering.html

Yes, you can use endmember spectra derived from other ROIs to perform a SAM classification because it can accept spectra and does not require the calculation of covariance like other algorithms such as Maximum likelihood. You can also use them for MF/MTMF or LSU. To do this, create a plot of the mean of the ROI and save it to a spectral library.

Hi,
Thank you for the reply. Should I always use SAM classification for the MTMF image? Or can I perform other classification methods such as SVM or Maximum likelihood classification for the MTMF image using ROIs I generated using my ground truth?
I used the hourglass to generate about 15 different endmembers (including the ones that are auto-generated by the software) and then finished the hourglass. Then I analysed the endmembers ROI file from the hourglass output and identified the endmembers that represent my land cover classes. Then I removed all the other endmembers and saved a new ROI file. I re-ran the hourglass using this new ROI file edited from the previous hourglass run. Still my results are very poor for the SAM classification. I tried SVM classification using my ground truth ROIs too. The results are better than the SAM, but still not satisfactory. Am I using the tool wrong? And is there any way I can make the endmembers better?

Thank you.

You cannot use a spectrum in classifiers such as max likelihood or SVM because these algorithms require the calculation of covariance, which requires as many samples as their are bands (nb+1). If you have ROIs with enough pixels, then you can use these other algorithms.

If you plot the mean spectrum of your classes, are they distinguishable from each other? If there is a lot of overlap (for example with different vegetation types, this is common), then it is usually difficult to get a very clean classification. if there are specific regions where they are distinguishable, you might try spectral feature fitting.
http://www.harrisgeospati...lFeatureFitting.html

I have lots of ROIs created from my ground truth data. I tried maximum likelihood classification for MTMF image created with WorldView 3 and it gave me good results. I am trying the same method with Sentinel 2 and getting poor results.
I have three vegetation classes with somewhat similar reflectance curves (trees, short grass species and a tall invasive grass). But they do have some regions that are not overlapping. I will try the spectral feature fitting. Thank you for the suggestion!