The segmentation method highly depends on how general it must be and the variability of your images.

What is rock, what are pebbles and what is sand anyway?

All those big circular contours are majorly the rocks. Pebbles are small ones of them. There is a threshold range between which I need to detect the contours and identify them a rocks.

Rest all plane sand is sand. For example circular contours with area between 10-1 units must be detected or with major axis length 4-8 units must be detected rest all ignored.

How about the variability of your images? The one you posted is not your own, but a stock image from 123RF.

Based on this image only, I ask myself: how do I discern the rocks? I think it is because they cast dark shadows.

I thresholded the red channel at 120, and applied a LocalMax 3x3 convolution filter (in order to remove small black spots), and did a distance transform: the further from a shadow, the higher the pixelvalue (see picture). In the picture you can see several rocks that are well identified, but also large sandy areas... You can get rid of the latter, by specifying a maximum size to the 'high distance transform areas'. In Matlab, the function ' bwconncomp ' will do the trick.

For your convenience (the distance transform is the Blue channel of your rock image).

Thanks a lot. Yea actually first I was testing on this image only... Can you please share the code so that I can get a more clarification about your processing.

Thanks in advance.

I programmed it with proprietary software from the Netherlands Cancer Institute. I described exactly what I did, so you could replicate it with Matlab or Octave.

Mind however that this method only works when rocks cast small shadows. Also the threshold is 'tricky'. I mean, other images probably need other threshold values.

Okay anyways thanks for the help!