Is Lee Sigma filter THE Lee Sigma?

Microwave Toolbox
#1

I’m working on an ENVISAT WSM data which have to be filtered. During my tests I applied both Lee Sigma and Refined Lee filters (from Speckle Filtering -> Single Product Speckle Filter).
Thanks to @jun_lu, I’ve found the appropriate references to the filters:


Lee et al. (2009) publication (an introduction of Improved Lee Sigma filter, in SNAP /and in this post/ Lee Sigma), presents a comparison of several filters, including Refined Lee and Lee Sigma. The comparison shows, that - in case of those two filters - the results are not so much different (Lee Sigma is slightly better). Both preserves edges, Lee Sigma resolves problem of blurreness, among others.

I applied:

  • Refined Lee
  • Lee Sigma (no. of looks: 1, window size: 5x5, sigma:0.5, target window size: 3x3)
  • Lee (thus, not Improved Lee filter, filter size: 3x3)

and others variations of the Lee/LeeSigma filters, which, for clarity of this post I’ll not mention. Despite of different parameters for the Lee Sigma, the result was similar or “worse” than presented here.

Surprisingly, LeeSigma caused the strongest “blurriness” of the SAR image. Based on Lee et al. (2009), this should not be the case. The Improved Lee Sigma filter should preserve edges and mitigate the problem of blurring which was the thing of Lee filter.

Therefore, my question is: Is the Lee Sigma really the Improved Lee Sigma speckle filter (Lee et al. 2009)?

The results of my tests:

image
image1162×579 104 KB

Just in case, reference to Lee et al. (2009):
Jong-Sen Lee, Jen-Hung Wen, T. L. Ainsworth, Kun-Shan Chen and A. J. Chen, “Improved Sigma Filter for Speckle Filtering of SAR Imagery,” in IEEE Transactions on Geoscience and Remote Sensing, vol. 47, no. 1, pp. 202-213, Jan. 2009, doi: 10.1109/TGRS.2008.2002881.

1 Like
#2

In [1], several filters have been compared including Lee Sigma and Improved Sigma (MMSE) filters. The “Lee Sigma” filter mentioned in the paper is the filter that was developed by Lee et.al.in 1983. The MMSE filter is the real Improved Lee Sigma filter that overcomes the bias estimation problem in the old “Lee Sigma” filter. The Improved Sigma (MMSE) filter is the one that we have implemented in SNAP as “Lee Sigma”. Maybe we should rename it “Improved Lee Sigma” to mitigate the confusion.

[1] Lee J S, Wen J H, Ainsworth T L, et al. Improved Sigma filter for speckle filtering of SAR imagery. IEEE Transactions on Geoscience Remote Sensing 2009; 47(1): 202-213.

1 Like
#3

@jun_lu, thank you for your quick answer and explanation of the naming.
I was not confused by the name, thanks to your posts which I’ve hyperlinked in my previous message (thank you for that!). I analysed results of Refined Lee and MMSE in [1] (e.g. Figure 5d and 5f). But definitely SNAP Help regarding speckle filtering could be complemented… (not necessary the renaming, just proper citation would be an asset).

In [1] results of Refined Lee and MMSE (Figure 5d and 5f; Figure 6d and 6f) show, that both filters preserves edges, there’s no blurriness. This is not the case of my results where MMSE (“Lee Sigma” on my picture) is definitely “the blurriest” filter… Based on [1] I would expect results comparable to “Refined Lee” and this makes me worried that the LeeSigma alghoritm is not the MMSE (i.e. a bug?..)…

[1] Lee et al. (2009) - cited in previous posts

1 Like
#4

Sorry I misunderstood you. We will look into it. Thanks for reporting the problem!
The problem has been tracked by ticket at https://senbox.atlassian.net/browse/SITBX-783

2 Likes
#5

Perfect, thanks a lot!

#7

Hi @atoco and @jun_lu,
do you know if this question has been solved in the last SNAP update? I have checked the link from @jun_lu but I haven’t seen any update about the ticket.
Thank you in advance.

S. Savastano

1 Like
#8

Hi @s.savastano, from time to time I have a look at the ticket and as you’ve noticed no updates so far… Maybe the ticket itself was not updated but the filter was? This I don’t know as I haven’t used it for some time.

#9

Dear @atoco,
thank you for your reply.
I hope @jun_lu or some other developers could help us on this matter.
In the meantime, I have a question for you. I am monitoring coastal erosion by S1 and I am trying to understand which speckle filter provided by SNAP is better for this analysis, preserving as much as possible the edge detection. I think you have a great experience about these filters and so do you have any advice to give me on this?
Thank you in advance for any tips or suggestions.

S. Savastano

#10

Hi guys, we have reviewed the algorithm and the code, and did not find anything wrong with the implementation. We have tested the operator with some ALOS product and the result is expected, i.e. the homogeneous area is filtered and edges and point targets are retained.

The reason that some filtered SAR image looks blurry could be because the image is over filtered due to incorrect number of looks. Basically the improved Lee sigma filter is an MMSE estimator. However when incorrect number of looks is used (say the actual nLooks is 2 and we used the default value 1), then the filter becomes a local mean filter which leads to the blurriness in the result. The information of the number of looks can be found in the abstracted metadata of the product.

If you still have problem, please share the product you used and the detailed processing steps. We will look into it. Thank you

2 Likes
#11

Thank you for testing and sharing these information.

Could you please give an example on this and how it is correctly used to adjust the filter?

#12

image
image1194×155 5.17 KB

1 Like
#13

thank you. So whatever is entered there (based on the processing of the product) is the best option for the Lee Sigma filter?

#14

Hi guys,
GRD data has got 5 range looks and 1 azimuth look. Which value has to be used in this case (if it should be necessary to apply a speckle filter on a GRD product)?

2 Likes
#15

Thank you @jun_lu and @ABraun to deal with this “issue” working with Lee Sigma filter.
Using this old question from @TDedring , I would like to ask if you have a response for it.
I am in the same situation like @TDedring . I am working with GRD product and I am applying the Lee Sigma filter but it’s not clear to me which number of Looks is suitable with this kind of data (from metadata 5 range looks and 1 azimuth look).
Could you kindly help me (us) on this?
Thank you in advance.

S Savastano

#16

S-1 GRD has ~4.3 equivalent looks (ENL):

IW GRD Resolutions - Sentinel-1 SAR Technical Guide - Sentinel Online - Sentinel Online (copernicus.eu)

2 Likes
#17

Thank you Marcus @mengdahl !!!
Always appreciated your help.

Best,
S Savastano

1 Like
#18

Dear Marcus,
I was looking at your suggestion but unfortunately there is something unclear to me.
You said to consider for GRD product a ENL around 4.3, but in the despleckle filter tab for Lee Sigma you can just select the Number of Looks (from 1 to 4) and not ENL.
Do you have any suggestion on this? Which number of Looks is more suitable for GRD product?
@ABraun @jun_lu have you some other tips on this matter?
Thank you in advance again.

S Savastano

#19

I would simply select 4 as it is close.

1 Like