Deconvolution and Regularization Demonstration - PowerPoint PPT Presentation

About This Presentation
Title:

Deconvolution and Regularization Demonstration

Description:

none – PowerPoint PPT presentation

Number of Views:90
Avg rating:3.0/5.0
Slides: 16
Provided by: ricka1
Learn more at: https://epsc.wustl.edu
Category:

less

Transcript and Presenter's Notes

Title: Deconvolution and Regularization Demonstration


1
Deconvolution and Regularization Demonstration
  • Rick Aster
  • CIG/EarthScope Imaging Workshop Tutorial
  • Washington University, St. Louis
  • 11/1/06

2
Goals
  • Demonstrate basic Matlab coding
  • Demonstrate Frequency-domain deconvolution
  • Demonstrate Time-domain devonvolution
  • Demonstrate deconvolution instability in the
    presence of noise
  • Illustrate fundamental smoothness vs. data fit
    tradeoffs with both freqeuncy-and time-domain
    regularization for the deconvolution inverse
    problem.

3
Here is the desired deconvolution result a
simple Earth impulse response.
4
Here is the convolving kernel (option 1) In the
case of an instrument response deconvolution,
this is typically a band-limited (smooth)
function which isthe recording system impulse
response. In the case of a receiver function it
is the vertical-comp seismogram.
5
The data is the convolution of the previous two
time function.
6
The spectral division solution is essentially
perfect if the data are noiseless and the
convolving kernel is known exactly
(double-precision Matlab default calculations).
7
Next, we add a small level of Gaussian white
noise to the data vector.
8
and all hell breaks loose!
9
We can see why the previous deconvolution result
is so terrible to deconvolve we divide the noisy
data spectrum by the convolving kernel spectrum
(smooth curve here)
10
thus producing a deconvolution for the noisy
data that is completely dominated by noise!
11
The water level regularization provides one way
to stabilize the deconvolution and select a best
result, based on fitting the data and having a
small model (deconvolution result) norm.
12
Here is a selected optimal regularized solution
selected using the previous tradeoff curve.
13
Tikhonov Regularization Tradeoff Curve
Same problem regularized via 2nd order Tikhonov
regularization
14
Here is an optimal model selected from the
previous tradeoff curve.
15
Deconvolution Demo Exploration
  • Noise level
  • Try a more complicated convolution kernel (e.g.,
    uncomment the optional convolution kernels and
    rerun the routine).
  • Try a different regularization scheme by altering
    the L roughening matrix in the time domain
    deconvolution.
Write a Comment
User Comments (0)
About PowerShow.com