Is it possible to apply the Otsu thresholding (or any other image processing algorithm dedicated to image digital values in the interval [0, 255]) to a Radiometrically calibrated Sentinel-1 SAR product for which the pixel values (sigma0) are signed real (negative or positive decimal) and are not normalized (logarithmic scale)?

If it's feasible, how should the intensity values be rescaled to the byte interval ([0, 255]) prior to the processing?


Otsu's method does not really care about actual values since it tries to minimize the total variance within classes while maximizing the distance between the classes. So, you could just run Otsu on your original data (no need to rescale) and it will provide you with the optimal threshold to use to achieve the goal listed above.

I don't know what is your main goal but, if you want to segment out specific regions, one of the problems with the original Otsu's method is that it is a global method and will treat your dataset as a series of points not spatially related to each other and, given the high dynamic range of speckle noise, it would probably be hard to clearly separate the two classes "foreground" from the "background".

I would suggest the use of a speckle reducing filter (a list of some can be found here) followed by a global or adaptive local thresholding method. This will provide you with the indexes of the pixel belonging to "foreground" and "background" and you can use these indexes to segment your original image, if this is your goal.

  • Thanks for your answer. Could you recommend any adaptive local thresholding method? – Hakim Oct 16 '15 at 20:30
  • 2
    This post has a couple of good references including optimization method to speed up the execution. Depending on what environment and/or libraries you are using, you might have at your disposal some of these methods. For example one common library that has ports for the most common packages (matlab, python, c/c++, etc) is OpenCV and it has a library for adaptive thresholding – Andrea Oct 17 '15 at 1:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.