Doris fine correlation parameters
Recently I noticed that some of my interferograms using TerraSAR-X data had "loss of coherence" in azimuth direction. I first thought that it might have been a problem with DORIS, but then it turned out to be a mistake in my settings.
The processing was done in ADORE with the default settings for fine correlation, which I had looked up from the DORIS user's manual. The default settings for the fine correlation in ADORE are:
ADORE: s fc_
fc_acc=4 4
fc_in_pos=
fc_initoff=coarsecorr
fc_method=magfft
fc_nwin=4000
fc_osfactor=16
fc_plot=0.4 NOBG
fc_winsize=32 32
You can see that I had already deviated from the DORIS user's manual and set the number of windows (nwin) to 4000 instead of 400. These settings however caused a "correlation loss" issue. After discussing with Mahmut Arikan from TU-DELFT, I was convinced that this was due to something going wrong in the coregistration of the slave image to master image.
Luckily, while I was processing an old dataset (Envisat, Mexico City) I noticed that one of the pairs that had been working fine had the correlation loss issue. This eventually led me to changing the fine correlation parameters. The working pair had the settings fc_osfactor=32, fc_winsize="64 64".
Test Results
I tested reprocessing a problematic scene with different oversampling factor (osfactor) and winsize. Below you can see the results of TerraSAR-X data acquired over Juarez, Mexico. The dates are 20100725 and 20100714 for master and slave. All interferograms are reference phase removed, showing topographic fringes with a perpendicular baseline of -117 meters.
osfactor=16,winsize="32 32",accuracy="4 4"
osfactor=32,winsize="32 32",accuracy="4 4"
There is little improvement with the increased oversampling factor. Then I thought I could increase the window size for better estimation of correlation between master and slave.
osfactor=32,winsize="64 64",accuracy="4 4"
Increasing the window size actually made the results worse. This actually makes sense, but you will have to look at the plots at the end of the text.
osfactor=32,winsize="16 16",accuracy="4 4"
And finally a good result...
Discussion
You may be wondering why increasing the oversampling factor and decreasing the window size helps the coregistration. DORIS basically calculates cross-correlation of many (4000) windows over master and slave images and uses the ones above threshold coherence. The windows are interpolated in the given oversampling factor, and since we need a maximum of 1/10th offset between pixels (of master and slave), increasing the oversampling factor, allows the offsets to be calculated more precisely.
The correlation values always decrease with larger windows, but it is also true that the very small window sizes will cause wrongly estimated high coherence values (Just and Bamler, 1994). Therefore reducing the window size increases the coherence leaving more points above threshold. Below you can see the outputs of dummy2.*.ps files for two cases. These plots show the offset vectors used in the coregistration parameters (coregpm) step.
dummy2*.ps, osfactor=32,winsize="64 64",accuracy="4 4"
dummy2*.ps, osfactor=32,winsize="16 16",accuracy="4 4"
But you can see that even though the number of offset vectors is increased, they still do not cover the full scene and lower azimuth line numbers still do not contribute any information. Looking at the DORIS user's manual again, we see that there are suggested values, which are different from the default values for several parameters for fine correlation step:
Recommended values are:
fc_winsize=64 64
fc_acc=8 8
fc_osfactor=32
Processing the interferogram with the recommended values provide a good result (best among the ones tested above). Below is the interferogram and the offset vectors:
osfactor=32,winsize="64 64",accuracy="8 8"
dummy2*.ps, osfactor=32,winsize="64 64",accuracy="8 8"
Conclusion
On June 1st, 2011 after our discussions with Mahmut Arikan (TU-Delft) we came to the conclusion that the problems were due to low accuracy settings (acc=4,4), which did not allow for coregistration of larger offsets. If speed is not an issue it may be best to set the accuracy to half the window size (acc=32,32), and use a large (64,64) setting for the window size. This would increase the robustness, with some loss in calculation speed. I'll be changing the default settings in ADORE accordingly, which will be available with revision 197.
References
|
