next up previous print clean
Next: Convergence issues Up: comparing IRLS and Huber Previous: Tests on synthetic data:

Tests on real data

Figure 8 shows the input for the real data example. These data are contaminated with a low velocity/low frequency noise crossing the main signal. The Huber threshold and the damping paramater for IRLS are equal and obey the empirical choice that we gave above. Figure 8 details the least-squares result. The velocity panel (middle) shows some horizontal stripes that severely blur the main velocity fairway; in addition, because no strong event appears below 1s., a poor resolution is achieved. In regard to the l2 result, we might note that the remodeled data (right panel of Figure 8) exhibit some artifacts in the top part that shows the limitations of the least-squares inversion in this case. Figure 10 displays the ``l1'' inversion results with the Huber solver and IRLS. Again, because we are able to pick up a velocity fairway all the way down, both results are very comparable. Moreover, the remodeled data (Figure 9) show fewer artifacts than they do in the l2 case.

 
gr-wz08-L2-HUBER
gr-wz08-L2-HUBER
Figure 8
From left to right: 1) Input data. 2) velocity domain, l2 inversion result. 3) remodeled data from the l2 result
view burn build edit restore

 
comp-data-wz08-0.082
comp-data-wz08-0.082
Figure 9
From left to right: 1) Remodeled data with l2. 2) Remodeled data with the Huber solver. 3) Remodeled data using IRLS
view burn build edit restore

 
comp-mod-wz08-0.082
comp-mod-wz08-0.082
Figure 10
From left to right: 1) l2 velocity space. 2) Huber velocity space. 3) IRLS velocity space
view burn build edit restore


next up previous print clean
Next: Convergence issues Up: comparing IRLS and Huber Previous: Tests on synthetic data:
Stanford Exploration Project
4/27/2000