Hi Alberto, You ask some good questions.  I'm afraid that the answer, though, is that it doesn't theoretically make sense to include more endmembers in your unmixing analysis than the number of dimensions of the data plus 1.  If your MNF analysis shows you that you have only three dimensions of data (i.e., three non-noise MNF bands), then you can only have a maximum of 4 endmembers in a valid unmixing analysis (and you only get 4 if you use constraints - without a constraint it's one less).  ENVI lets you choose as many endmembers as you like, but that doesn't mean that the results are valid.  You mention that your RMS error image has very low values, which is a good sign.  Does it contain any recognizable features?  That would indicate a problem in the analysis. I'm also guessing that you used constrained unmixing, which gives you abundance images with reasonable values, even if the model is not valid.  I recommend that you try completely unconstrained unmixing.  That way, if your set of endmembers do not comprise a valid model for your data (and if you have 5 endmembers with a dataset that only has 3 real dimensions, then your model is not valid for your data), that will be reflected in unreasonable endmember abundances.   If you'd like to create images of abundance without the endmembers being limited, you might want to try ENVI's Mixture Tuned Match Filter tool.  It calculates each abundance image for each endmember without reference to any of the other endmembers.  So, in that case, you can choose as many endmembers as you wish.  It still doesn't mean that you really can distinguish that many endmembers from your data, but each individual result should be as valid as possible.  Keep in mind, in MTMF, that each endmember you choose needs to be rare in the scene. I hope this helps. Peg

Hi Peg,   Thanks for the reply, it was really helpful. I´ve tried many different ways of running the process: constrained, unconstrained… but not all of the abundance images have reasonable values. The RMS images have recognizable features, so you must be right, and the model is not valid for my data. I guess it means that I can not apply the Linear Spectral Unimixing on my data, right?   Could I try the 4 MNF components (actually the third and the fourth got similar eigenvalues) with the 5 endmembers in a constrained unimixing analysis, would that be correct? In that case, how can I know which weight factor should I use?   I´ll try the MTMF just with those endmembers representing sealed areas, which are just three.   Thanks again   Alberto

Hi Alberto, Yeah, when you can't get reasonable values in your abundance images, especially with unconstrained unmixing, then that's a good clue that your endmembers don't constitute a good model for your data.  And that seems to be confirmed by the real features you see in your RMS error image.  If you need to determine the abundances of your five selected endmembers within each pixel of your scene, then I would conclude that Linear Spectral Unmixing doesn't seem to be a good analysis tool for your purposes.  If you can be flexible about the endmembers you use, then you should be able to get reasonable results using unmixing, *if* you are able to find the right set of endmembers.  To do that, you might want to experiment with ENVI's SMACC tool (Spectral > SMACC Endmember Extraction). Personally, I don't recommend using constrained unmixing at all.  It can make your endmember abundances look resonable, when the endmembers themselves are not appropriate.  For an extreme example, consider what would happen if you had, say, a sine wave as an endmember.  You'd get reasonable results, sure.  But a sine wave is not a real endmember for your scene.  So, those reasonable-looking results are totally wrong. My point is that constraining unmixing confuses the interpretation of the analysis.  If you've defined your spectral endmembers well, then you should get reasonable endmember abundances without constraining the unmixing.  I hope that helps. Peg