Topic: Motion detection of Lambert Glacier

I am a postgrad student doing a research project by using COSI-Corr and ASTER imagery to detect the motion of Lambert Glacier in Antarctica. As we don't have an available DEM so have to use the DEM extraction function of ENVI to generate the DEM from 3N and 3B bands.

The generated shaded DEM is messy and has a lot of TIN looking due to the shadow of rocks and the high contrast of ice and snow cover. So I only can roughly pick up tie points between the shaded DEM and the first raw ASTER image as I can hardly identify the areas in shaded DEM. The optimization of GCPs either failed because of 'array has too many elements'(what does it mean?) or has a lot of misregistration (NaN). Can you please let me know by what means I could probably have a better registration result such as how big the window size I can use?

Or can I ID a point in the raw image and then locate that point in the shaded DEM when I am choosing the tie points?

I understand that a good DEM is very important for COSI-Corr. But the DEM I have so far is the best I can use.

Thank you very much for your time and the help.


Re: Motion detection of Lambert Glacier


if you are barely able to identify similar areas in your raw image and in you shaded DEM, there is no way the correlator will.
The GCPS optimization failure is due to the DEM quality. 'Array has too many elements' is an error happening typically during the resampling and is caused by a DEM containing outliers that corrupted the orthorectification process; the misregistration (NaN) are due to unsuccessful correlation between the shaded DEM and the ortho Aster image.

If the 3B/3N ASTER is your only source of DEM, I would suggest you try to obtain the best DEM possible by testing the various options available during the DEM extraction. In particular enlarge the correlation chips size to maximize the chance of good correlation. Once you get the best DEM possible, have a close look at it (using a shaded version of it):

- If it looks good, without a lot of TIN pattern, and you are able to recognize areas between the shaded DEM and the Aster image, then pickup you tie-points again, and try a GCPS optimization. Ususally a window size of at least 128 (256 best) is recommended. Be careful however that the DEM is clear of outliers in the area covered by the 128 or 256 pixel chip centered on you tie-points. If the GCPS Optimization still fails because of a too smooth topography (leading to a shaded DEM without enough texture for the correlator to succeed), then select your tie-points as precisely as possible between the shaded DEM and Aster image, but do not optimize them (you will do the optimization only between the first orthorectified Aster image and the second Aster image).
Also, as the DEM will still probably contains some outliers, verify the resampling distance (see 8.2 and 10.6 in the COSI-Corr doc for more details), and enter them manually if necessary.

- If you still see a lot of TIN patterns and you are still unable to locate areas with your first ASTER image then do not use it. It is sometimes better to not use a DEM than using a really noisy one. Give it a try without DEM. Depending on the glacier you are studying, it sometimes happens that the area of interest is on a relatively flat topography (in your case that will correspond to a glacier that is flowing on a gentle slope). In that case, the effect of the topography will be negligible on your area of interest. When selecting your tie-points between the first ortho Aster image and the second Aster image, try to select points with the same altitude (a coastline, if by chance you have one in your images, will be helpful).

So, to sum up:
Get the best DEM possible, then successively try (depending on failure or success):
- GCPS optimization
- Tie-points selection, without optimization
- No DEM use

Do not hesitate if you have any other questions/issues.



Re: Motion detection of Lambert Glacier

Dear Shirley,

I addition to Francois's post, let me add a few comments. It is true that DEM quality may have a strong impact on the quality of the correlation results, but the influence of the DEM also depends on the incidence angle difference between your images. In short, if you're working with images having small incidence angle difference (<1 degree), you will still be able to obtain good results, although the DEM may be of poor quality. As a rule of thumb, ASTER images with almost exact footprint overlap should have very similar incidence angles.

If optimization of GCP between the shaded DEM and the image fails because of poor DEM quality, you have two possibilities:
1- select a few tie points very carefully and do not optimize them (the good old times method, Francois solution #2)
2- Do not select any tie points and orthorectify your ASTER image without using any GCP. This solution should work fine if you didn't use any GCP when generating your DEM.

As Francois wrote, it is true that you could not use any DEM at all and see what happens (but most likely the results will be impossible to interpret).

Try to use a DEM whenever possible, even if this means not optimizing the GCP, or not selecting any tie-points. Whenever you use a DEM, to make sure that nothing weird happens during processing, the DEM should not contain any "missing values" or unrealistic values. When DEM generation fails, missing values are sometimes replaced with -65535 values, or Nan, etc... Any such values will make your processing to fail, as COSI-Corr will interpret these values as elevations of -65535 meters. You should interpolate the missing values to be within a physical range, although "it might not look good".

For your project, I would then advise:
1- Try to use images with almost identical footprint to minimize potential parallax error due to poor DEM
2- Generate a DEM and make sure it doesn't contain any missing value (interpolate the missing values if needed)
3- If DEM generation didn't use any GCP, do not use GCP either to ortho-rectify your ASTER image, it should give acceptable results. If results are not acceptable (you can check a posteriori if the ortho-image and the shaded DEM overlap properly), you can use tie points, without optimizing them if the shaded DEM visually looks bad.
4- Regardless of the previous steps, you will have to select and optimize GCPs (maybe 5-6) between the second raw image of your project, and the ortho-image generated in step #3.

Hope this help, let us know what worked best for you.



Re: Motion detection of Lambert Glacier

Hi Shirley,

yes, correct, I forgot to mention the possibility of working with images having a small incidence angle difference. That is a valid option too. Thanks Seb !!

Regarding suggestion #3 from Sebastien's post, some comments to avoid a classical trap:
You can orthorectify the first ASTER image without taking any GCP as long as the DEM was extracted without GCP AND was extracted with the 3B and 3N bands of the same acquisition (of your image). So for example, if you have two ASTER acquisitions (let's name them 1 and 2), and you extract the DEM (without GCP) with 3B and 3N of acquisition 2, you MUST orthorectify first the image from acquisition 2. Then you can go on with the classical procedure: tie-points selection between the orthorectified image 2 and raw image 1, GCPS optimization, image 1 orthorectification using the (same) DEM.
The reason is that without GCP involved, the georeferencing is defined with the satellite ancillary data only. You therefore need to use the same acquisition for the image to be orthorectified and the DEM extraction in order to have coherent georeferencing between them.



Re: Motion detection of Lambert Glacier

Hi Sebastien and Francois,

Firstly thank you very much for your help. I finally got the correlation images without using any GCPs during the first orthorectification/resampling process. They look good! And the velocity detected is similar with the previous studies.

But I have some problems when trying to analysis the images.

1) The velocity distribution over the Mawson Escarpment (the rocks surface) are uneven and the ice seems to flow to every direction.  What is the reason?

2) There are noises and decorrelation on the main ice stream. I cannot tell which are caused by topographic features (high contrast from ice and snow) and which are from satellite effects such as CCD array? Do you need me to email you the images?

Thank you very much.




Re: Motion detection of Lambert Glacier

Hello Shirley,

Glad to read that it's progressing...

1) I don't know the Mawson escarpment, but it sounds like a topographically rough area. If it is the case and if you used the ASTER DEM to orthorectify your image, it is most likely that the artifact you're seeing are due to topography residual. The ASTER DEM, which has a probable real resolution of around 100m, does not have a sufficient resolution to correct for all topography of the 15m ASTER image. The uncorrected topography is therefore retrieved in the correlation maps. To reduce these artifact, you either need to have access to a better DEM or find a pair of images with smaller difference of incidence angle (see Seb's post #1)).

2) With ASTER imagery, there is no obvious CCD artifact as far as we know. The main artefact that is usually encountered with ASTER is its perturbated attitude that is retrieved in the correlation maps as an undulating pattern running in the along-track direction. The noise and decorrelation are mostly due to temporal decorrelation and in that case it is recommended to increase the correlation window size.



Re: Motion detection of Lambert Glacier

Hi Shirley,

Depending on the quality of your images, interpretation of your data can be difficult. The first thing to check is to make sure that you used the best parameters for the correlation:
1- window size: look at the maximum displacement to be expected in your images. Glaciers can flow quite fast and depending on the time separating the 2 images, the displacement to be measured can indeed be large. You should always use correlation windows that are at least 2 or 3 times as large.  Say that you expect ~200m of displacement. If your pixel size is 15m, you should use window sizes of 200/15*2 = 26.6 pixels.  This is quite close to a 32 pixel window, then to be "safe", the multiscale version of the correlation should be used. The larger window should be large enough (here you could use 64x64 pixel windows), and the smaller window can be 32x32. you can use a step of 8 or 4 pixels, depending on the resolution you want in the displacement field. Use 4 robustness iterations.
2- After the correlation, make sure to discard larger outliers. You can do this easily by using the COSI-Corr -> tools -> discard/replace image values. Discard all the values outside the range of expected measurement, and with SNR less than 0.9.
3- When you have your 2 ortho-images, open then both in different displays and link them (image window -> tool -> link -> link display, then click ok). When clicking on the images, you will go back and forth between the 2 images. Inspect your images for changes. As a rule of thumb, if, at the scale of the correlation window you cannot recognize the two images, correlation will fail.
4- You can try to build a vector field Tools -> Vector field, then export the vector field as an ENVI Layer. You can then super-impose the vector field to the ortho-images (you will be asked the coordinate system when you export the vector field, it has to match the one from your images). Then you can check that the values are not too inconsistent. Try to identify why the correlation did not work. Ice surface dramatically changed? Changes in shadows orientation? New vegetation? Clouds? Change in snow cover? Images are saturated because the ice/snow is too bright? Using ASTER images, if some vectors are consistently oriented to much along the East-West direction, this could mean that the DEM you are using is not of very good quality or that significant ice thicness change occured during the acquisitions.

Hope this help.



Re: Motion detection of Lambert Glacier

Hello Shirley,

maybe you can use de RAMP DEM of Antarctica

http://nsidc.org/data/docs/daac/nsidc00 … v2.gd.html

or now the ASTERGDEM:


Best wishes,