In very homogeneous areas (in terms of backscatter), you can find an offset by chance (false positive). The way SNAP corrects that is by low pass filtering the results.
It exists better ways to filter offset tracking results. I would suggest this method : https://www.mdpi.com/2072-4292/9/10/1062
The algorithm of SNAP itself is quite common and you will find in the literature accuracy results of offset-based measurements depending of your field of interest (glaciology, earthquake, landslide, …).
Thanks. That is we don’t assess reliability of displacement (velocity) in terms of probability / likelihood at all. We assume that velocity field is ok in general and we just need to remove some outliers…
Where can I find low pass filtering in SNAP (interface)?
In the parameter of the Offset Tracking tool, you can see that SNAP is performing a spatial average, ticked by default with 5x5 box size. That can be considered as a very simple low pass filtering
I strongly recommend to assess at least at known position your results. Maybe you have land-surveyor GPS data or maybe a place you know it doesn’t move so that you can compare SNAP results with a ground truth.
Many artifacts can alter pixel offset (PO) results. Change in the ionosphere can produce azimuthal strikes and alter PO measurements, just to give one example
It is now clear. But why not to resolve reliability issue by simple increase in cross-correlation threshold? Let say up to 80%. Yes, a number of calculated GCP (and the size of velocity field) will sharply decrease. But any case a low maximum cross-correlation value, e.g.0,2, shows no real match.
You should not mentally scale the correlation as a direct 0->100% match between your master and you slave image. If the correlation is statistically different from 0, a match occured.
0.2 threshold is already quite high and in low backscatter, you’ll miss opportunity to find good matches by increasing too much the threshold.
I recommend first doing so trials and check if it makes sense in areas you have info (like stationary areas).
Hi Sina, did you find this publication?
No Unfortunately, In this community as far as I understood the focus is on high level usage of the products and not on the underlying methodology. In general the offset tracking algorithms are implemented based on either nomalized cross correlation of the image with respect to a defined kernel (window), or based on calculation of optical flow to calculate the motion maps and the disparity maps. But as you might have noticed it is not quite clear how did they implemented it.