A new method for more efficient seismic image segmentation

Accuracy comparison

Of course, computational efficiency means little if the resulting segmentation is not accurate. For the synthetic image result in Figure 7(a), we can get a qualitative sense for the accuracy of the new segmentation implementation; while the salt body is divided into multiple segments, these segments do indeed fit almost exactly inside the boundaries of the salt body. A more direct comparison of the relative accuracy of the NCIS and PRC methods can be obtained via analysis of the field data results. Figure 9 shows both calculated salt boundaries overlain on the image: the NCIS boundary in green, and the PRC boundary in pink. Visually, we can see very little difference between these two results; in many locations, they are almost exactly on top of one another. The most noticeable difference between the two results is near $X=20000 m$ , where the PRC boundary dips deeper than the NCIS boundary. Examination of the input image in Figure 1(b) suggests that in this location, an error in the migration velocity model has led to a discontinuity in the boundary image. The new method appears to do a better job of ``correcting'' the error. This result serves to increase confidence in the viability of this new segmentation scheme as an alternative to the existing NCIS implementation.

uno-segcomp
Figure 9. Comparison of the boundaries obtained using the NCIS eigenvector method (green) and the pairwise region comparison method (pink).

A new method for more efficient seismic image segmentation