Hi Mona, Regarding the mask, it sounds like maybe you applied the mask to the image, converting all of the pixels under the mask to the same pixel values? If so, then yes, it is normal that all of those pixels would be given the same abundance in your unmixing output. Once you have applied a mask (versus specifying the mask during input to the unmixing tool), ENVI doesn't retain information about where the mask was applied, so it doesn't know that those pixels are masked. It treats them just as if they were valid pixels with the assigned values. It is curious that you say you did not find any pure pixels using the Pixel Purity Index. I'm not sure how that could happen. The algorithm should be able to identify pure pixels in any multispectral image. Unmixing with only three bands usually doesn't give great results. This is because you have only 2 degrees of freedom, so the algorithm can only distinguish 2 endmembers. It could be that the endmembers that are distinguishable with the bands in the data are not materials that are of particular interest to you. They are likely to be vegetation and some aggregate non-vegetation endmember. The way you chose your endmembers may not be the most valid way to find endmembers in the image. If you are looking at only 2 of your MNF bands, then there is a third MNF dimension that is not being accounted for at all. I am guessing that is why your resulting fractions are going beyond 1 and below 0. If those abundance values are common in the scene, then it is telling you that a linear unmixing model using your endmember spectra does not model the actual spectra in the input image very well. A better way to find your three endmembers, if for some reason PPI doesn't work on these data (although it certainly should) may be to create a ROI that covers most of the scene, and use that as input to the n-Dimensional Visualizer. Then you can choose the most extreme pixels in the three corners of the data cloud that you see in that tool as your endmember spectra. In this situation, where you have very little spectral information (only 3 bands, which I assume are visible bands and therefore highly correlated), you may very well find better results using mixture tuned matched filtering. That is a valid choice. Regards, Peg Exelis VIS