Large holes in CARTOSAT-1 DEM using AMES in hilly terrain

[Arnab Sarkar]

Hi, I have been trying to generate a DEM using CARTOSAT-1 stereo pairs.

I am following the following steps:

1) Bundle adjusting the RPCs. I am not using any external GCPs.

2) Using Bundle adjusted RPCs to run the parallel_stereo command. The command I am using is 
parallel_stereo AFT.tif FWD.tif AFT.adjusted_rpc.xml FWD.adjusted_rpc.xml run/run --alignment-method local_epipolar --corr-kernel 5 5 --subpixel-mode 3 --corr-max-levels 5 --corr-memory-limit-mb 12000 --stereo-algorithm asp_mgm --nodata-value 0 --corr-seed-mode 1 --sgm-collar-size 128 --threads 8

3) point2dem to convert point cloud to DEM with 3 m resolution.

But in hilly terrains, I am having a lot of voids. Can you please suggest some changes that can improve the quality? Also, it will be a great help if you could suggest some techniques to which files I should look to optimize my parameters in the parallel stereo command.

Regards,

Arnab

[Oleg Alexandrov]

You can try mapprojecting the input images. That should help with hilly terrain. See https://stereopipeline.readthedocs.io/en/latest/next_steps.html#mapproj-example

It is very important to get a DEM relative to the ellipsoid, as discussed. Then mapproject the left and right image with the same local projection (also discussed) with same grid size which is close to the native ground sample distance for your case.

You should inspect the left and right mapprojected images and see if there is any notable large shift. There can be some local shits. Then, the features on the input DEM, such as hills, should agree well with their locations in mapprojected images. Any notable large scale shift or large local shift (over 150 pixel say) can result in slow processing or suggests the DEM is not consitent with the inputs. You can overlay the mapprojected images and input DEM in stereo_gui as georeferenced images from the menu.

As to suggestions, stereo-algorithm asp_mgm with subpixel mode 9 is likely good enough.

During correlation, look for the search range entry in the log files or printed to terminal. If it is say of width of height over 150 pixels that means it will take time.

Also note that the asp_mgm algorithm can use a lot of memory, so need to set careffully number of processes per node depending on what you have on your machine. Then can monitor cpu and memory usage as it runs.

--
You received this message because you are subscribed to the Google Groups "Ames Stereo Pipeline Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ames-stereo-pipeline...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ames-stereo-pipeline-support/1ba1f53d-6d31-4562-a365-5a29894b3ff8n%40googlegroups.com.

[Arnab Sarkar]

Thanks for your quick response.

I have two queries:

1) Do u suggest doing bundle adjustment before map protection and using the bundle adjusted rpcs for map projection?

2) If my reference dem is of 30 m. Should I map-project CARTOSAT images to 2.5 m or 30 m?

Best regards,

Arnab

[Oleg Alexandrov]

I have two queries:

1) Do u suggest doing bundle adjustment before map protection and using the bundle adjusted rpcs for map projection?

Sure you can. Bundle adjustment will not create new RPC. it will create .adjust files that you can pass to mapproject and parallel_stereo with --bundle-adjust-prefix. The doc for this tool has examples.

After you run this, look in the output directory for initial and final residual stats files. Those will tell you initial and final pixel reprojection per camera. If it changed very little, likely bundle adjustment is not necessary, but usually it is suggested.

 

2) If my reference dem is of 30 m. Should I map-project CARTOSAT images to 2.5 m or 30 m?

You should mapproject at the image resolution. The DEM will be interpolated to that. If you don't set mapproject resolution for one image it will be auto-determined, then you can use the same for the second one (the mapproject doc has an example for how to reuse a mapproj image grid size and extent). But it is better if you have from the vendor or in metadata the image resolution (ground sample distance, grid size) to use that one.

 

As before, inspect after mapproject the input images and DEM by overlay. That is most important predictor. Then look in bundle adjust output dir for stereo convergence angle. That should be 20-40 degrees hopefully.

To view this discussion visit https://groups.google.com/d/msgid/ames-stereo-pipeline-support/0a0f3433-999f-4d4d-8293-86c411c08a67n%40googlegroups.com.

[Shashank Bhushan]

Hi Arnab,

To add on the mapprojection step as suggested by Oleg, gridding the output DEM to >2 to 4 times the input image resolution will also reduce the gaps. For our research applications we have gridded DEMs from Cartosat-1 at either 5 m or 10 m resolution.

Cheers,

Shashank

To view this discussion visit https://groups.google.com/d/msgid/ames-stereo-pipeline-support/CADDkNmMfvzA9TMu6meynkGP4hWfkTtNUM%2BM19XRTfT72NdbX%2Bg%40mail.gmail.com.

[Arnab Sarkar]

Thank you all for the key insights you shared. I will try to incorporate all these suggestions.