If the PPI plot doesn't level out into a horizontal line, then it means you are still finding new extreme pixels, and you need to continue with more iterations.  If it seems like you are finding too many, then you can reduce the threshold, and see if that helps.  But, probably, you just need to do more iterations. Apparently, you have not found them all in 30000 iterations, and more are needed to find them all.  If you don't wait until you've found all of the extreme pixels, then you will likely miss some endmembers when you go to the n-d visualizer step.  If you miss some endmembers, then your linear unmixing model will not be valid.  So, it's pretty important to keep going until your curve flattens out, if you plan to do linear unmixing. Peg

50.000 iterations with a threshold of 2,5 still don't seem to be enough. Is it possible that this is a result of incomplete data preprocessing or a faulty FLAASH reflectance conversion?

Try 100,000 iterations with a threshold of 1.0.  I suppose it could be the result of bad data, but I'm not sure how that would lead to problems getting PPI to finish.  Do the data look good?  Are the reflectance values in the FLAASH result reasonable?  It might be helpful to do some exploring of those data, for example by calculating statistics, and checking to see that the stats seem reasonable.

Somehow I am still not able to get a proper result. So I've been thinking what are the factors of influence: 1. reflectance file 2. mnf transformation 3. PPI settings: 3a. spectral subset 3b. spatial subset 3c. threshold 3d. iterations 3e. memory or file So lets assume that that FLAASH and MNF worked properly. 3a. spectral subset: I usually try to utilize those mnf bands with eigenvalue >1. I don't want to loose information by just utilizing the first ~20 bands with high eigenvalue. On the other hand I'm not sure what will happen if I would use all bands, including those with mainly noise... 3b. I don't know if I am off the track here by doing a PPI calculation on just a spatial subset. Don't think thats the right approach as the scene has to be seen as a whole (depending on the task) 3c. Im not sure here how the threshold influences the fact that not all pure pixels are found. A threshold of 1 still doesn't help. 3d. Is it possible that 100.000 iterations still are not enough? 3e. Something odd I noticed is that fact that there is a difference if I save a PPI calculation to file or put it into memory. When putting it into memory after every 250 iterations many more pixels are found than with the s a m e settings saved as a file. Doesn't make sense to me. Putting it into memory roughly 1500 pixels are found every 250 iterations, but only 150-300 when saving it as a file. The same symptome occurs when I work with a spatial subset or the whole scene. I'm beginning to think that maybe I did something wrong during FLAASH or MNF. I used the Hyperion Tools from Devin to read the dataset into ENVI. I then changed the bad bands list back from 155 bands to 196 (only calibrated ones minus overlapping). I then started FLAASH (without wavelenght recalibration and spectral polishing - I'm bifid whether those 2 options are useful or not and if selecting them or not will change my results). The results seem to be okay as far as I can say. Although, the water vapor image and cloud mask seem somehow strange to me. To me the dataset seems faulty judging the output... I uploaded (klick) an extract for someone to judge. There is some kind of haze in the scene which you can spot in a RGB depiction (in the lower right corner). Nevertheless those vertical lines through the whole scene make me worry. For the MNF Transformation I used a water surfaze for the noise stat estimation. MNF band Help please.

Hi Cirendl, This is the first ever time I have joined the ittvis community. These days I am doing the spectral unmixing of Landsat TM data and I am facing the same problems as you mentioned in your posts. I just want to know if you could find answers to your queries. If yes, can you please help me about them? Thanks in advance.