1

Topic: statistical correlator

I am comparing results from IMCORR and COSI-Corr's statistical correlator from a pair of Landsat 5 images. The two give different results using similar parameters. Via the published literature I understand quite well how IMCORR works, but I have found nothing describing how COSI-Corr's statistical correlator works, so I cannot explain why they give differing results. Can anyone tell me if COSI-Corr's statistical correlator uses the same algorithm as IMCORR, and/or if COSI-Corr calculates the peak of the correlation differently than IMCORR? This may help explain the differences in the results between the two feature tracking programs.

Thanks in advance,

Jamie

2

Re: statistical correlator

Dear Jamie,

COSI-Corr calculates the correlation peak simply using the absolute value of the correlation coefficient formula. Then the correlation peak is interpolated using a quadratic polynomial for sub-pixel accuracy. We haven't put much effort into the statistical correlator, it's mostly provided as a sanity check or if images are very noisy. The frequency correlator should give you better results in general.

Sebastien

3

Re: statistical correlator

Dear Sebastien,

Could you let me know what is the oversampling factor used in the statistical correlator? I need this parameter in my analysis.

Thanks.
Luyi

4

Re: statistical correlator

Dear Luyi,

The statistical correlator does not use oversampling.

Sebastien

5 (edited by LuyiSun 2016-04-06 07:58:03)

Re: statistical correlator

Sebastien_Leprince wrote:

Dear Luyi,

The statistical correlator does not use oversampling.

Sebastien

Dear Sebastien,

Sorry I misunderstood the procedure. As you mentioned in this post "Then the correlation peak is interpolated using a quadratic polynomial for sub-pixel accuracy" , I wonder what accuracy (fraction of pixel) it can achieve in the Cross Correlation?

Thanks,
Luyi

6

Re: statistical correlator

The statistical correlator wasn't really optimized for subpixel accuracy and the simple interpolation we performed gives an accuracy which is usually around 1/4-1/3 of the pixel size. It does show some staircase biases at the subpixel level. The frequency correlation was optimized for subpixel accuracy.

Sebastien