pygmtsar
pygmtsar copied to clipboard
AlexeyPechnikov
[Help]: Question on Gastein Valley and reference points
I have replicated the Gastein Valley notebook on an area with known subsidence. The issue is in cumulative displacement, the stable structures appear as uplift, the known susidence location appears as almost no displacement, and another location nearby with extreme subsidence appears as subsiding. Thus, we have a bias towards uplift that can only be corrected through reference point selection.
I have a couple of questions:
- Regarding Gastein Valley, how do you know there is not a bias in the displacement results? Did you correlated your timeseries with GNSS data? Or was it a randomly selected "stable point" compared to another "subsidence point" that may or may not be actually subsiding.
- What do you think is the cause of known bias towards uplift or subsidence in sbas results even when detrending?
1 - Relatively stable area (red)
2 - Known subsidence area (green)
3 - Known extreme subsidence area (blue)
For the Gastein Valley, I used some ground data to identify stable and moving points. While these may not be highly accurate, they are sufficient for comparison. To validate processing stability, we can compare different correlation thresholds (and resolutions). The analysis of Gastein Valley results shows that for correlation thresholds of 0.3 and 0.5, the RMSE is less than 0.05, aligning closely with the threshold of 0.7 which yields the lowest RMSE. However, for a threshold of 0.25, the RMSE increases to 0.06, comparable to the no-threshold calculation with an RMSE of 0.084. These findings suggest that the most likely scenario is where the POI0 point remains stable.
What do you think is the cause of known bias towards uplift or subsidence in sbas results even when detrending?
The bias in SBAS results towards uplift or subsidence, even after detrending, can significantly depend on the chosen detrending method. Gaussian detrending, though simple and effective for SBAS resolutions of 60-90 meters, may struggle to distinguish between small atmospheric fluctuations and the actual ground signal. A more robust method is regression, used to remove topography-related delays and linear trends, particularly during the topography phase removal step. An example illustrates the difference between SNAPHU unwrapped multi-looking phase with trend removal and a recalculated single-look phase with trend removal during the topography correction step.
Please note, the Google Colab (not Pro) example for PyGMTSAR is limited to only the best scenes and interferograms. While this may yield optimal results, it's not suitable for statistical validation due to the insufficient range of interferograms. For more complex analysis, Google Colab Pro or local processing is necessary. These more detailed examples are not shared on GitHub.
The POI0 point is close to the 3 pixels with velocities -1, -1.5, -3 mm/year defined on this map: https://egms.land.copernicus.eu/#llh=13.18248411,47.24009181,2500.95535719&look=-0.15478280,-0.73398728,-0.66129038&right=0.98747286,-0.09417063,-0.12660667&up=-0.03065355,0.67260283,-0.73936851&layers=VHR%20Image%20Mosaic%202018_VHR%20Image%20Mosaic%202018-VHR_2018_WM:None,L3-U-U2-release&points=L3-U-U2-release,45881615696149602 There are also some corner reflectors in this area.
For the Gastein Valley, I used some ground data to identify stable and moving points. While these may not be highly accurate, they are sufficient for comparison. To validate processing stability, we can compare different correlation thresholds (and resolutions). The analysis of Gastein Valley results shows that for correlation thresholds of 0.3 and 0.5, the RMSE is less than 0.05, aligning closely with the threshold of 0.7 which yields the lowest RMSE. However, for a threshold of 0.25, the RMSE increases to 0.06, comparable to the no-threshold calculation with an RMSE of 0.084. These findings suggest that the most likely scenario is where the POI0 point remains stable.
Are you referring to the SBAS coherence mask or the PS coherence mask when you compare these 0.7, 0,.5, 0.3, 0.25 thresholds?
For the single-pixel plots mentioned above, I am referring to the PS correlation mask in the time series.
For the single-pixel plots mentioned above, I am referring to the PS correlation mask in the time series.
But the correlation mask given in the notebook example is the mean correlation used to discard arbitrary non-PS points based on mean correlation of the entire interferogram collection. In the plots above, you seem to use a different approach of removing pairs based on pair correlation and not mean correlation of the entire stack.
I share my notebooks with different approaches. At the end of the day, all the methods should produce the same result. And AI PyGMTSAR assistant even can help you to find the differences :)
For the Gastein Valley, I used some ground data to identify stable and moving points. While these may not be highly accurate, they are sufficient for comparison. To validate processing stability, we can compare different correlation thresholds (and resolutions). The analysis of Gastein Valley results shows that for correlation thresholds of 0.3 and 0.5, the RMSE is less than 0.05, aligning closely with the threshold of 0.7 which yields the lowest RMSE. However, for a threshold of 0.25, the RMSE increases to 0.06, comparable to the no-threshold calculation with an RMSE of 0.084. These findings suggest that the most likely scenario is where the POI0 point remains stable.
Alexey, which command can I use to reproduce this baseline displacement?
The command to plot is included into the notebook at https//InSAR.dev