1- The mask threshold in the correlation helps to discard frequency components with low power. The phase of low power components is mostly random and it's usually beneficial to discard it. The lower the threshold, the more frequency components are discarded. A value of 0.9 should work well in almost all cases.
2- You can use the correlation tool with the multiscale approach: the correlation window should be at least twice as large as the largest displacement to be measured. However, when you enlarge the correlation window, you are going to lose some spatial resolution and the displacement field will be more blurry. To avoid this problem, you can use several window sizes, for example 64 down to 32, or 128 down to 32, etc... COSI-Corr first correlates with the larger windows, then re-center the correlation according to the first result and recomputes the re-centered correlation with a smaller window, until the smallest specified window is reached. Therefore you can measure large displacements and still maintain the spatial resolution given by smaller windows. For your case, I would assume that a correlation with windows 64 -> 32 pixels should be optimum. Do not use windows smaller than 32x32 pixels, results will become meaningless with all 8 bits quantized images (e.g., ASTER, SPOT, etc...).
3- Striping artifacts usually have an amplitude around +/- 10m. If glacier displacements are larger than that, the striping will be saturated and the images won't have enough dynamic to show the stripes. To check for stripes or waves artifacts, you have to crop all the values that are outside the +/-10m or +/-15m range (in short, you have to discard most of the signal from the glaciers). You can use the "discard value" tool if you want to save the results, or you can adjust the display "Enhance -> interactive stretching" in the ENVI display, and reduce the range, then click apply.
4- In the destriping tool, there is a field "image to define correction from". This is the file you're going to input to determine the destriping model. In your case, the destriping is a simple averaging of the residual displacements along the across-track of the sensor. Assume you have glaciers in the correlation with large displacements. You certainly don't want to average the signal of the glacier to build the destriping model. Thus, you can build a correlation file where glaciers are masked out. An easy solution is to discard all measurements that may represent some signal by cropping large values. This is what is represented in Fig. 4 of Scherler's paper (note that glaciers appear white because they have been cropped out). Then COSI-Corr determines the global bias, and this global bias will be subtracted from the original measurements. Therefore, you should input the original correlation file in the field "image to apply correction". Note that you don't always need to differentiate between the image used to derive the correction, and the image to which the correction is applied, depending on how your glaciers (or other processes) are distributed in your image. For another example of destriping, you can look at Fig 1 of this paper:
http://www.tectonics.caltech.edu/slip_h … SS2007.pdf
Hope things make more sense. It's a very easy process, but a little tedious to explain, sorry:)