If the polygons in your screenshot are the typical size of acres in the aea, I think the filtering technique is of minor importance because you probably average the backscatter over the whole polygon. I personally would select an edge-preserving filter (Lee Sigma) or the IDAN filter (region growing) which level out extreme values but still maintain the edges between the different crops.
Small window sizes should be better at your scale.
Hi Stefano, if you are interested in the backscatter values within each polygon and their variation within an individual image and between images taken on different dates, personally I would not use any filters. I do not see the value of changing the statistics of the pixels within each polygon before calculating the radar cross-section.
In fact, I do not have to investigate the variability of the values within a single polygon (for example with a cross-section), but I have to investigate the time series trend of the average value of all the pixels inside a polygon (… many polygons), in a period that starts from April 1st to October 30th.
I am using GRDH data all in the same relative orbit and all in acending mode in north east Italy.
The polygons were created with a object-based segmentation technique on a Sentinel-2 cloud free scene, so they are all homogeneous within them (homogeneous from an agricultural point of view).
If you only need the average value within a field there is no need to speckle-filter - just compute the average directly
They might be quasi-homogeneous from an optical remote sensing viewpoint, but I would not call that “agricultural point of view”. You can also segment an average radar backscatter image of the same area and get quasi-homogeneous segments from a radar RS viewpoint.