Hello,
I have been processing a set of 5 TerraSAR-X Stripmap SSC images (HH polarization, same beam, all descending, acquired July–September 2023). I am using ESA SNAP 13 with the snaphu plugin for phase unwrapping. I have several questions and would be very grateful for any guidance.
Processing chain so far
For each pair I followed this chain:
- Co-registration of the two SLC images
- Subset to AOI (approximately 14,213 x 16,994 pixels at ~2m resolution)
- Interferogram formation (flat earth phase subtracted, topographic phase NOT subtracted, targeting a full DSM rather than differential heights)
- Goldstein adaptive filter (alpha 0.2, FFT size 32, window 3, coherence threshold 0.3)
- No multilooking applied as I don’t want to lose spatial resolution
- Snaphu export and unwrapping
- Snaphu import
- Range Doppler Terrain Correction on the unwrapped phase band (Copernicus 30m, 2m pixel spacing, UTM 22S)
- Export as GeoTIFF
Issue 1 — Tile boundary artifacts (chessboard pattern) in snaphu
I have been running snaphu with TOPO cost mode, MCF init method, 10x10 tiles, ROWOVRLP/COLOVRLP 400, NPROC 4, TILECOSTTHRESH 500. Several of my pairs show a clear chessboard pattern of squares in the unwrapped phase. The squares appear to align with tile boundaries.
I have tried:
- Increasing overlap from 200 to 400 pixels
- Reducing NPROC from 8 to 4
- Reducing tiles to 5x5, caused RAM crash
Some pairs with lower coherence consistently produce the Exceeded maximum number of secondary arcs error during tile assembly, even when all 100 tiles complete successfully.
My questions:
- What is the recommended approach to eliminate tile boundary artifacts without exceeding available RAM?
- Could coherence masking before snaphu export fix for low-coherence scenes?
Issue 2 — Phase to Elevation smoothing over buildings
When I run Phase to Elevation (using Copernicus 30m as reference DEM) followed by Range Doppler Terrain Correction, the resulting DSM looks smooth and does not resolve individual buildings, even though buildings are clearly visible in both the filtered wrapped interferogram and the unwrapped phase. Tall isolated buildings that are visible in the coherence image and interferogram are completely absent from the final DSM.
I am including screenshots of a zoomed-in and full extent of (a) the filtered wrapped interferogram, (b) the unwrapped phase over the same urban area for reference, and (c) the final DSM produced with phase to height and then range-doppler corrected, compared to the LiDAR DSM reference.
(a)
(b)
(c)
My hypothesis is that Phase to Elevation uses the reference DEM to anchor the phase-to-height conversion, and since Copernicus 30m is too coarse, the operator effectively pulls pixel values toward the smooth reference surface. Is this understanding correct? If so, is there a way to reduce this effect? For example by masking building footprints during the spatial correlation step?
Issue 3 — Alternative approach: multi-pair phase regression
To bypass the Phase to Elevation smoothing, I implemented a pixel-wise linear regression approach inspired by Weissgerger et al. (2017, Remote Sensing, doi:10.3390/rs9101010), which derives height directly from the relationship φ = -kz·h across multiple interferometric pairs, without using a reference DEM in the conversion step. I computed kz for each pair from the orbital parameters in the snaphu.conf files, applied a flat-ground phase offset correction, and ran the regression across 3 geocoded unwrapped phase GeoTIFFs that did not have the chessboard pattern artifacts.
The result shows a much better correspondence with a LiDAR reference DSM, and critically, isolated tall buildings that were absent from the Phase to Elevation result are now visible. However the result is noisier than the Phase to Elevation output, and buildings do not align perfectly with the LiDAR due to the side-looking geometry of TSX (layover, foreshortening).
I am attaching a comparison map between the regression result and the LiDAR reference.
My questions:
- Does the Phase to Elevation operator in SNAP indeed smooth results toward the reference DEM, or is there another explanation for the building signal loss?
- Are there better approaches within SNAP to perform the phase-to-height conversion without this smoothing effect?
Issue 4 — Building height estimation from InSAR or other methods
More broadly, I would welcome any suggestions from the community on methods to estimate building heights from TerraSAR-X Stripmap data, whether interferometric or otherwise. I am aware of PSI approaches but have only 5 images, which is insufficient for a full PSI stack. Are there other methods that could work with a small number of TSX Stripmap acquisitions over a dense urban area?
Thank you very much for any guidance. I am happy to share additional screenshots or processing parameters if helpful!
Guilherme