Classification of GRD product

Hello everyone,

I would like to realize a supervised classification using sentinel 1A (GRD) images i did this steps with snap : radiometric calibration
speckle filtering
terrain correction ,
now i’m looking about steps to classifay my image? thank you for your help!

1 Like

If you only have VV polarization classification is difficult. Typical supervised classification is based on samples which help to divide your feature space. But even if you have VV and VH, your feature space is only n=2 so your classes won’t be much distinctive. To increase your feature space you have the following options:

  • add image textures (GLCM module), but this should be done before speckle filtering. Speckle filtering destroys most of the image texture.
  • add images of different dates (dry and rainy season, for example)
  • add images of a different sensor
  • add topographic information (right-click > add elevation band)

As soon as you have a stack of bands with more than about 4 bands, supervised classification makes more sense. But if you are using inputs of different units (textures) you can no longer use the Maximum Likelihood or Minimum Distance classifiers. I’d recommend the Random Forest classifier.

I classified image based on SAR textures and additional bands for classifications and it worked out well:
http://www.int-arch-photogramm-remote-sens-spatial-inf-sci.net/XL-7-W3/777/2015/
http://hw.oeaw.ac.at/?arp=0x00324a97

If you have specific questions on one of the single steps, feel free to ask.

Related topics collected here.

7 Likes

thank you for your speed answer, yes it very helpful.
so I’m beginner with snap and sentinel 1.
My product GRD content to polaraisation HH AND VV .
i added GLCM and elevation.
but when i would classify my image I dont know how to realize this steps with snap.
really I am blocked.
I need methodology .
thank you for you help.

This is explained here (the example is based on Sentinel-2 but it also works with a SAR data, textures and elevation)
I have never tried it but in the meanwhile they do not even need to belong to the same stack but can be added as different products in the classifier.

You can start based on this and ask for specific steps when you stumble upon problems. At best with a screenshot so we can help more effectively.

my question now how can i choose the region of interest and my image is black?
its so different about sentinel 2 .
what is the role of the backscattering?

it should not be black, does it contain data?

You should know that if you want to work with radar data :slight_smile: It looks different because SAR data are different than multispectral data.
Maybe you first read some basics:
http://www2.geog.ucl.ac.uk/~mdisney/teaching/PPRS/PPRS_7/esa_sar_tutorial.pdf


http://www.eamnet.eu/cms/sites/eamnet.eu/files/Introduction_to_SAR-Marjolaine_Rouault.pdf

1 Like

Sorry , my image with grayscale

this is how radar data typically look like :slight_smile:

1 Like

my problem is how to class this images , and how can i identify the classes ( vegetation , water …)

there is a tutorial “interpreting SAR images” here:
http://www.crisp.nus.edu.sg/~research/tutorial/sar_int.htm

there is a box called “how to read SAR images” here:
https://www.asf.alaska.edu/about-sar/sar-basics/

there are also nice lessons on SAR backscatter here:
https://saredu.dlr.de/unit#category=*&type=lesson

Why starting with something complex such as classification without some image basics? :slight_smile:

5 Likes

Hello , @ABraun
I have a problem with executing PCA ,I tried to follow the steps of the tutorial
http://www.int-arch-photogramm-remote-sens-spatial-inf-sci.net/XL-7-W3/777/2015/isprsarchives-XL-7-W3-777-2015.pdf that you offered to me , i would like to execute PCA,The program runs without stopping

the input :

PCA can take quite much time. Try 8 components instead of -1 or skip it and directly input all layers in the classification.

Looks good by the way :slight_smile:

I have to do something similar so I need to know if i get it rigth,:

Callibrate then GLCM, then with the GLCM do the filtering and the terrain correction?
and with the original callibrate images also do the filtering and the terrain correction?

so I will have all the GLCM and plus the 2 sigma 0vv and sigma0 vh?

Yes, makes sense.

Andreas is right about applying GLCM before filtering but Keep in mind that S1 is C-SAR with a moderate resolution. If in your future work dealing with high SAR resolution like COSMO-SkyMed I suggest you to apply filter before GLCM in order to get a result.

Here is a nice PHD dissertation about it.

http://www.disp.uniroma2.it/geoinformation/students/geoinformation-dissertations/Pratola_PhD.pdf

Actually, I wouldn’t filter at all. If you are aiming at the image’s texture, filtering before GLCM destroys much of it.

1 Like

I meant after. that would be OK or not. but the terrain correction must be done in any case , as far as I understand no?

On more think I have read your paper http://www.int-arch-photogramm-remote-sens-spatial-inf-sci.net/XL-7-W3/777/2015/isprsarchives-XL-7-W3-777-2015.pdf . There you did not filter (as suggested) did terrain correction and also run glcm for textures. My question is did you do the terrain correction before or after glcm. I believe it was after. But I want to be sure.

Calculating GLCM befofe Terrain Correction should be better because the intensities are not resampled yet.
Filtering the textures is not prohibited. If it enhances the information content, wyx not :slight_smile: