Phase discontinuities at burst boundaries when stitching multiple burst interferograms

[AntariaQAQ]

Hello @scottstanie and the Dolphin team,
Thank you for developing and maintaining Dolphin. I am using it for time-series processing of a single Sentinel-1 frame, and I would greatly appreciate your advice regarding an issue with burst stitching.

Processing information

• Sentinel-1 ascending, relative orbit 40, OPERA Frame 127, VV polarization
• 20 acquisitions from August 4, 2024, to March 20, 2025
• 27 bursts across IW1, IW2, and IW3
• Sentinel-1 SLCs processed into Geo-CSLCs using COMPASS 0.5.7
• Dolphin version: 0.42.7 development build, commit 2cbe461
• Output resolution: approximately 30 m

Issue

Based on the intermediate outputs, the issue appears to occur when multiple burst-wise interferograms are stitched into a single frame-wide interferogram, before phase unwrapping begins.
The attached figure shows the frame-wide interferogram for the 20241003_20241015 pair. It was generated directly from the phase of the stitched complex .int.tif file. I observed clear phase discontinuities along several burst boundaries.
The relevant Dolphin step appears to be the following call in stitching_bursts.py:
date_to_ifg_path = stitching.merge_by_date(...)
As far as I understand, merge_by_date() groups the burst interferograms by date and passes them to merge_images(), where gdal_merge.py is used to create the stitched raster.

Questions

Is this type of burst-boundary phase discontinuity commonly encountered when processing multiple bursts with Dolphin?
Among the 54 interferogram pairs generated in this run, approximately five show seams as clearly as the attached example. These discontinuities also affect the subsequent phase-unwrapping results.
Do you have any recommendations for diagnosing or correcting this issue? I have been considering whether the phase information in the overlapping areas between adjacent bursts could be used to estimate their relative phase offsets and apply a phase-leveling correction before stitching. Would this be a meaningful approach, or would you recommend a different solution?
If helpful, I would be happy to provide the Dolphin configuration, individual burst interferograms, or additional overlap statistics.
Thank you very much for your time and assistance.

[scottstanie]

Depending on the input processing, yes this could be reasonably common, as some areas have enough ionosphere that geocoded SLCs do show burst ramps.

I have been considering whether the phase information in the overlapping areas between adjacent bursts could be used to estimate their relative phase offsets and apply a phase-leveling correction before stitching

I also thought that was worth trying at some point, and it is on my development branch here:
scottstanie@ec458ac

you can try using that/porting it to your own if you'd like. I haven't done extensive testing on it, just saw a few improvements in cases like this

[scottstanie]

i agree that the constant seems good as the default to me, and i'm not sure when you'd need/want the high order right now. also agree that i thinnk getting rid of the large jumps should be enough to fix phase unwrapping!

[renierlgv]

Agree. Still, seems like the correction operates on the AOI, which doesn’t work well in small areas with low coherence. If I’m correct in my interpretation of how the correction works it would be good to change it to use the entire overlapping area for the offset estimation regardless of the AOI.

[scottstanie]

hmm that's a good question (though more of a sweets than a dolphin question), as I can't recall if sweets crops the ifgs before stitching, during, or after. i believe even if you have a small aoi, you make GSLCs that are the full burst size though.

[renierlgv]

Right. On the dolphin side, it is possible that spurt could introduce a bias while converting the single reference network into the 3 nearest so adjacent bursts that have no phase inconsistency still produce a seam ? This was the output of a ai analysis I did on a problematic case but can’t recall further details. Not sure if this is of any help at all

[scottstanie]

I don't think that sounds right to me. i could see a small jump turning into a full cycle slip during unwrapping, but i dont think "introduce a bias while converting the single reference network into the 3 nearest" is right, as after phase linking, the redundant network is just for ease of unwrapping and introduces no new phase

[renierlgv]

an unrelated question @AntariaQAQ, will be too much to ask how did phase linking go for you on this huge area?

[AntariaQAQ]

To be honest, my understanding is that Dolphin’s phase linking is essentially performed on a burst-by-burst basis. In terms of spatial coverage, it mainly depends on how many bursts are mosaicked together.

For a single burst, I have tested the performance of phase linking before, and in some cases the coherence improvement could reach up to around 50%, so the effect can be quite noticeable. However, for areas with dense fringes, if the window size is not chosen appropriately, there is also a possibility that some originally clear fringes may become slightly blurred.

So I think it would be better to evaluate its performance based on the specific characteristics of your dataset and the actual situation during processing.

[renierlgv]

True, yet I also believe that the mask of your AOI applies here. Do you have some benchmarks of how much time it takes the step for you (one full burst)? what is the resolution of your GSLS 5x10? A full burst will take me 2 weeks or more, so, I just would like to have a real case scenario to compare time. I need to figure out if resource management is my problem.