power-point talk
Download
Report
Transcript power-point talk
Image Estimation by Example
(Claerbout’s winter class)
and its relation to Andre’s
recent tomography lecture
Large Linear Problems
Large= lots of data, Model= image
Linear= too much data for nonlinear
Why? Each survey $5M and 5 terabytes.
Subtract two surveys for fluid movement and permeability barriors.
Two-stage least squares
1. Find a differential equation that converts a
training image to random noise.
(will be the “regularization operator”)
2. Where the data fails to determine the
model, solve the differential equation
(This is “regularization”.)
Jump to movies
vector vector
= vector
- vector
0 residual = theoretical data
- observed data
0
m
-
d
matrix vector
-
vector
(image)
-
0
0
L
(5-D space)(2-D space) -
(television signal)
(3-D space)
L is an operator, not a matrix!!
• Matrix: a table of numbers
• Operator: two programs
– Prog1 = L m
– Prog2 = L’ r
(L’ is L transpose)
If you are using matrices, you are doomed!
Use Krylov subspace methods (conjugate gradients)
0 1
0
0
0 0
0 0
0
1
0
0
0
0
1
0
0m1 7
0m2 3
0m3 4
1m4 2
0 1
0
0
0 0
0 0
0
1
0
0
0
0
1
0
0m1 7
0m2 3
0m3 4
1m4 2
0 0
0
0
0 0
0 0
0
1
0
0
0
0
0
0
0m1 7
0m2 3
0m3 4
1m4 2
0 0
0
0
0 0
0 0
0
1
0
0
0
0
0
0
0m1 7
0m2 3
0m3 4
1m4 2
The null space has two members.
1
0
0
0
0
0
1
0
You can add any amount of these two
vectors to the solution m without
changing the theoretical data.
L is a matrix from physics
Example: line integral between two wells.
L usually has a null space.
0
0
e
L
m - d
A
m
(fitting)
(regularization)
(noise) (tiny) (statistics)
Guess A, or find A by classical autoregression,
also called a PEF=Prediction-Error-Filter
0 L d
m
0 eA n
d=data m=model (image)
L= operator from physics
A= model covariance destructor
epsilon=tiny number
n= optional noise, often taken zero
Finding the GEOFIZ solution and the GEOSTAT solution
Let n= random white noise
0
0
e
L
m - d
(data fitting)
A
m - n
(regularization)
GEOFIZ: Find m taking n=0
GEOSTAT: Find many m’s, one for each random noise sample
Thesis titles
• Reservoir simulation
• Reservoir uncertainty
• Reservoir estimation