1 (edited by seven 2010-09-11 06:49:18)

Topic: H Parameter

Hi Sebastien,

I am looking at denoising the E/W and N/S correlation bands through the COSI-Corr NL Means Filtering procedure.

My current displacements from sand dunes are roughly in the order of 16m over 5 years from two ASTER images, derived from a 64-32 window size with 4 pixel step.

I just have a question regarding the noise parameter, H. As I understand it a good approximation for H is 1.6 * std dev of the noise. The manual discusses obtaining the noise from local statistics of the image, but I am slightly confused as to which values I am measuring exactly to obtain the noise std dev?

Many thanks.


Re: H Parameter

Hi seven,

You can use the ENVI tool ROI, select and area of stable ground, and look at the local statistics. The stdev will give you an approximation of the noise level. For your case, the h parameter should be around 1-2. You can do trial and errors on small subsets.


3 (edited by seven 2010-09-14 17:41:19)

Re: H Parameter

Thanks for your help again Sebastien!

Just to clarify,

The displacement bands I have at the moment have had low SNR removed and erroneously high displacements cropped. I took the st dev statistics of noise from stable ground before after I did this, I then ran the NL means filter from the H value derived from this (estimated at 1.5 from the st dev). The st dev between the noise is reduced by approx 30%. Is this a sound method?



Re: H Parameter

You can increase the H value to reduce the noise further, but then you might blur the results too much, it's a compromise. Another solution, but it really increases the computation time, is to increase the search area.



Re: H Parameter

Hello Sebastien,

i have a question about the H-Parameter. I am working with Landsat-8-scenes of South America and i'm trying to figure out the correct H-parameter. As a method to assess the H-Parameter, you described a Estimation with the Stdv of noise. I am slightly confused, because in my case the stdv of noise in the Correlation-SNR-Band for a stable underground is far below one (e.g. 0,003).

If i used your hint (1,6*stdv noise) the H-Parameter would be also far below one. Shouldn't i use the stdv of noise of the SNR-Band or what else is my error in reasoning?
Thanks for help.



Re: H Parameter


You should only filter the EW and NS bands, it doesn't make sense to filter the SNR band.



Re: H Parameter

Oh, there is a misunderstanding. I don't filter the SNR-Band.

For the estimation of H, should i use the stdv of EW/NS-displacement of a stable ground or the stdv of noise for the same area?


Re: H Parameter

Hi Erik,

you should use the stdv of EW/NS-displacement on an area that is stable ground with no (or very few) outliers. The goal is to get the stdv of the natural noise of the correlation (not the outliers). Make sure to take an area large enough. Sufficiently large to represent the measurement of several independant correlation windows. For instance if correlation window size is 32 and step 4, then don't take a 8x8 patch to compute your sdtdv. Get at least 2-3 times larger than that.