H-alpha Wishart unsupervised classification does not match H-alpha plane

I have several quad-pol images which I would like to classify by H-alpha Wishart segmentation and analyse the scattering properties of several features (covered by mask) based on H-alpha plane.

For that, I performed unsupervised H-alpha Wishart Quad classification, after Terrain Correction I’ve added mask of my features and generated histogram for my mask. The classes in a mask are: 2,4,5,6,7,8.
Parameters of the classification: 5 window size, 100 iterations (based on my experience the more iterations the more accurate results but please, correct me if that’s not the case here).

In addition, I’ve performed H-A-alpha Decomposition. I’ve generated the H-alpha plane for my mask. The classes which contain pixels of my mask are: 3, 4, 5, 6, 7, 8 (below).


The density of H-alpha plane shows that class 6 should be the most frequent in a mask. However, the histogram shows that the most frequent is class 4 of H-a Wishart segmentation. The representation of other classes is also inconsistent when H-alpha plane and results of H-a Wishart classification are compared. Why such big inconsistency?

EDIT: I have also generated H-alpha Wishart classification for default parameters, i.e. 5 window size, 3 iterations. Here is a histogram for the same mask as above:


Indeed, class nr 6 is the most frequent (as on H-alpha plane) but there is also quite frequent class nr 2 (not represented by H-alpha plane at all!). There is no representation of class nr 3 (present on H-alpha plane).
Again: could you please explain me why this inconsistency?
(I suspect that there is misclassification of class 3 as class 2 in H-alpha Wishart classification process…)
How many iterations should be used for H-alpha Wishart classification? Just by comparing the histograms, there is a big difference between the results of 100 iterations and 3 iterations (not to mention the differences in comparison to H-alpha plane…). As I wrote before, I was taught the more iterations the better… I don’t really care about the computation time (dependent on iteration number) as long as I get correct results.

What is workflow (steps) do you use and what it is your target?
I’m searching for analysis of my data for land cover classifications and identification of buried subsurface features.

I need your advice

Thank you

As far as I remember I followed the workflow proposed in PolSARPro tutorials and ESA courses related to this software. E.g.:
http://seom.esa.int/polarimetrycourse2015/files/PolSAR_Practical_EPottier.pdf - at a slide 112 is an illustration of a workflow (and on the next slides their description).

So, I’ve adapted this workflow to SNAP.

Some links related to the PolSARPro courses which may help you:



and SNAP tutorial:

1 Like