Yes, got your data, thanks.
In theory, the maximum displacement that can be measured is 1/2 the correlation window. Therefore, in theory, if you have a 32x32 pixel window, you can measure up to 16 pixels of displacement, or 16*15=240m with ASTER. In practice however, this bound is a lot lower due to noise, multi-temporal changes, etc. Therefore, rather than hoping for a 1/2 ratio, we're usually happy when we can attain a 1/5 ratio. Meaning measuring ~5-6 pixels of displacement with a 32x32 pixel window.
It's the reason we implemented the multi-size windows. Often times, you want to be able to measure larger displacement, but still would like the advantage of smaller windows in terms of local averaging. Therefore we start with a larger window to get the large first order displacement, then we refine with a smaller window. That's what the 64-32 or 128-32 means. We do 2 passes, the first pass with a larger window, the second with a smaller one, re-adjusted according to the previous measurement.
For your case, you have to acknowledge that your images are not ideal. For the images separated by less than 1 year, you shouldn't expect a large displacement, therefore 32x32 pixel windows should suffice. However, the images from the same year aren't from the same season and you can notice a lot of differences in snow cover. That clearly constitutes noise that will damage the correlation.
On your image pair from 2001-2006, there are other problems. First, note that the main glacier is covered by clouds in some areas, so you can't expect the correlation to work there. On another area, you see the strong shadow cast from the clouds. You can't correlate there either.
Then, there are more problems: the displacement to be measured is quite large with respect to the width of the glacier. Remember, to measure large displacements, you have to use large windows. The problem, is that at the resolution of the ASTER image, a 64 pixel window will likely be larger than the width of the glacier. The correlation therefore becomes more difficult as the displacement is not coherent anymore within each correlation window.
You can see that on the faster parts, the correlation tends to give results going uphill on the glacier. This makes no sense, and it simply means that the texture on the glacier completely disappeared and cannot be traced anymore. By chance, the correlation finds other structures on the glacier that looks like what was there before, but it's obviously wrong.
My guess for these images: use a 64-32 scheme for the longer term pair, and 32-32 for the shorter term image pair, with a step of 2 pixels and robustness of 4 pixels. Then apply a directional filter, in the same fashion as Scherler et al (you'll have to do that, say in Matlab or IDL, etc.). Meaning you will manually define a direction for the glaciers you are interested in (eg central flow line), and you discard displacement vectors that strongly disagree with this a priori direction of flow.
You would get much better results if you could find images without the clouds on the glaciers, and with no more than 3-4 years between pairs.