Do you know what timescale / frequency range the target reflections occur? If so, try doing a Fourier transform of each IMF and work out which ones cover the frequency of the event. It might be simpler to add these together rather than to subtract the rest from the signal.

If it occurs at a low-frequency (e.g. 7th or 8th IMF) then I would remove all of the IMFs that are generated before it from the original signal, although this won't remove any of the low-frequency noise.

It is best not to run EMD to high-order IMFs as mixing between the modes results in large errors that then propagate through the decomposition. Also consider using ensemble EMD [*] where you add different noise realisations to the signal and find the average IMF (this can help reduce the mixing effect).

Good luck!

[*] Z. Wu, N. E. Huang, Ensemble empirical mode decomposition: a noise-assisted data analysis method, Adv. Adapt. Data Anal. 1 (1) (2009) 1-41

Dear Vahid,

One good way to decide is based on the amplitude spectra of the original data and the one corresponding to each IMFs. I suggest you take a look on the attached papers with my students.

Milton

Conference Paper Atenuação do ground-roll utilizando os métodos de deconvoluç...

You can see this paper for example of using EMD

Conference Paper A visual based approach for drowsiness detection

Dear Vahid,

It is possible that you select the most relevant IMFs according to a recent paper presented at ICASSP 2014 [1], and then apply the statistical significance analysis on the EMD, as mentioned in [2], fig. 3.11 or fig.3.12 (for signal contaminated with fractional Gaussian noise) or [2], chapter 5 (for signal contaminated with WGN).

[1] Douglas Baptista de Souza, Jocelyn Chanussot, Anne-Catherine Favre, ON SELECTING RELEVANT INTRINSIC MODE FUNCTIONS IN EMPIRICAL MODE

DECOMPOSITION: AN ENERGY-BASED APPROACH, 2014 IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP).

[2] N. Huang, S. Shen (eds.), Hilbert-Huang Transform and Its Applications,

http://www.worldscientific.com/doi/pdf/10.1142/9789812703347_fmatter

> The main idea is that non-stationarity caused by noise is collected in just a few IMFs that do not carry the target (primary) reflections in the seismic dataset.

Did you determine the correlation between the IMFs, and estimating the lag that gives the maximum value? Such lag may give you information on the reflection you mentioned.

Regards,

Fernando

I am grateful to all. I'm working on it and if I progress, I will let you know.

Ok, keep me updated please. Although I am experienced in EMD and related tools, I always find new interesting applications. It is possible that the next one is yours!

It may be VERY helpful this webpage:

http://rcada.ncu.edu.tw/research1.htm

Dear Vahid Sarli,

for signal denoising you can also use the band-variable filter. For any question do not esitate to contact me.

Sincerely,

Rocco Ditommaso

Article Analysis of non-stationary structural systems by using a ban...

Dear Vahid, 

You can use hurst exponent/ detrended fluctuation analysis to calculate which IMFs to keep and which to discard. The code for the same can be found here. 

http://www.mathworks.com/matlabcentral/fileexchange/52502-denoising-signals-using-empirical-mode-decomposition-and-hurst-analysis

Best wishes