Discriminate water surface and desert with Sentinel-1

Dear all,
I am trying to understand how to distinguish between water and deserts in Sentinel-1 SAR images.
Due to the very low backscatter intensity from both land cover types, I end up with a ton of false positive occurrences over deserts (desert areas classified as water).

I wasn’t able to find satisfying scientific literature on such topic. Can anyone suggest me some possible techniques?
Many thanks,

You can use a multi-temporal set, deserts will be much more stable over time than water. Or explore InSAR-coherence, which is essentially zero over water.

Thanks a lot for your suggestion. Indeed, I am currently using these features:

  • 4 mean images, each one obtained by averaging all the available images within 3 months (so, 4 mean composites for a single year).
  • A minimum composite on the whole year.
  • A maximum composite on the whole year.
  • A mean composite on the whole year.
  • A standard deviation on the whole year.
    So in total I have 8 images to train my classifier (16 if I use both VV and VH channels).

However, even with this choice of features, I wasn’t able to discriminate properly the two classes. I will investigate also InSAR-coherence.

Depending on the type of desert coherence might be close to zero over the dark areas too. Have you tried Principal Components? Also if you are in an area where the waterline is stable it might make sense to simply treat it separately from the land classes using a water mask in GIS.

I didn’t try to use PC. But I will also investigate this. Also, can you explain exactly what you mean with “treat it separately”? Thank you very much!

Well, for example you can mask water out, classify all the land classes (desert classification should work better), and then add the water-class back by using the mask.

Ah, yes. Trying to focus on deserts is something I didn’t think about. I’ll definetly do some tries and see what happens. Again, thank you!