understanding aps estimations parameters and results
Tagged: APS estimation
January 30, 2018 at 1:33 pm #2643
firstly we would like to congratulate you for the great work that has been put in the sarproz software.
We would like to get a better understanding on analyzing the parameters and results of the aps processing. We are running 66 sentinel-1 images, from 2014 to 2018, over an area of nearly 16×16 km using a star graph.
In the sparse point selection we are running the Amp. Stab. Index 1-Sigma/Mu > 0.72 and not 0.8 due to the low number of scatterers as you can see. In the graph creation we have selected local redundant (best) with 150 max r due to the way the scatterers are placed, the fact the the area covers high grounds, mountains and plains, as we believe its better this way to be analyzed.
Moving on in the proccesing parameters we are running the smart model with 5, knowing that the phenomenon is seasonal subsidence so the trend must be non-linear.
As we understand, the height parameter regards which will be the max and min height residual we expect, so we have been running some test with different meters (100, 80 , 60), this one is with a threshold of -30 and 30 meters. The results are as shown in the images.
Firstly we would like to ask you how can we understand which high and low thresholds are the appropriate ones for the analysis in the height parameters and what should we be looking in the results?
Should we look at the sigma connections residual heights to check if the are closer to 0? or should we see if the connections residual heights are spread appropriatly?
Adnd secondly we would like to understand what how we check if the aps estimates are the appropriate ones to move on to the sparse point processing.
We have have uploaded images on the results of this analysis, so you can have an understanding of the area and its circumstances.
If you need anything else to upload please tell us.
Thank you again and always congrats on the work.
January 30, 2018 at 1:37 pm #2648
January 30, 2018 at 2:12 pm #2650
The result looks reasonable.
To check if the APS was well estimated, you the final “test” function. If the final coherence is high and evenly distributed, you are done. You can also plot the APS one by one to check for issues.
To understand if the height estimation range is correct, check the histogram of the connections residual height. You can notice peaks at +-30. That could be noise, but it suggests that maybe a range +-40m is more appropriate.
In the past we built several graph types that have different performances depending on the case at hand. but recently we have introduced a new module to analyze and improve a graph after a first estimation round: http://www.sarproz.com/groups/sarproz-group/forum/topic/sarproz-release-2017-5/ If you want to try this module, use a delaunay graph first. And then run that module
January 30, 2018 at 3:25 pm #2651
Thank you for the super fast response.
we are running now the data with what you suggested.
Regarding the graph analysis and refinement, we have looked at it, but we do not understand the parameters exactly what they exactly they are for. Checking the manual it doesn’t really go in depth:
3. press the button “go” in the “clusters processing” frame to apply the chosen threshold. The software will tell you how many clusters are generated
4. plot the clusters with the “plot” button
5. choose a radius to select local references within clusters and press the “go” button next to it
6. view local references with the “plot” button next to the radius.
7. choose the method to adopt (a. connecting clusters via the local references, b. processing only the main cluster, c. processing clusters separately)
8. you can optionally add more points to the graph using the “points densification” frame: choose a parameter, add a threshold, choose a maximum distance
(w.r.t. the existing points), choose whether to use or skip the initial mask, press the button “go”
9. you can optionally display the new points and the new graph with the corresponding buttons
would it be possible to have a better understanding on these selections: a. connecting clusters via the local references, b. processing only the main cluster, c. processing clusters separately
and on what threshold should it be wise to select on the “clusters processing”, because it automatically selects the threshold from the non-linear weighting of the aps processing, is it ok to continue with this one?
Thank you again,
January 30, 2018 at 10:26 pm #2654
About a. b. and c. just go with a. The other options are still experimental.
About thresholds, use the suggested ones as starting point. If you are not happy, try to rise the minimum value for being more restrictive.
February 1, 2018 at 3:27 pm #2664
I apologise for the delayed thank you.
We appreciate the help. The results, with carefully consideration of the above suggestions, were found really close to the expectation we had.
Thank you again and always congrats on the work being done.
August 21, 2019 at 3:25 am #4576
I have a question about APS estimation. I work with 19 TerraSAR-X images in HH polarization. For APS estimation, I have been running some test with different threshold of height estimation ([-5 5],[-10 10],[-15 15], [-25 25], [-30 30],[-35 35]). The results are as shown in the attached word.
I would like to ask you how can we understand which thresholds are the appropriate ones for the analysis in the height parameters?
If sigma connections residual heights are closer to 0, they will be appropriate?
How can we understand which thresholds are the appropriate ones for the analysis in the linear trend parameters?
August 21, 2019 at 3:41 am #4577
August 21, 2019 at 3:42 am #4582
These are another attached file.
August 22, 2019 at 6:54 am #4586
usually, the best approach is to keep a quite large set of range taking into account the expected values by considering:
– the characteristics of your AOI (e.g. presence of high rise buildings);
– the type of graph you created (long connections, short connections, etc. )
Once you have a preliminary estimation, you can have a look at the histograms (as you correctly did) in order to be sure that the distribution is fine. To be more clear, in the -5+5 estimation you have too many connections estimated in the extreme bins and this is not good since it means that higher and lower values exists and they are probably not properly estimated.
Still the same condition is present for the -10+10 (see the peaks at the limit values). The 25 an 30 values are much more better. Keep much larger ranges will not lead you to a better result.
You must be logged in to reply to this topic.