A new method for more efficient seismic image segmentation

Conclusions

This paper presented an implementation of the pairwise region comparison (PRC) scheme of Felzenszwalb and Huttenlocher (2004) for segmenting seismic images. Numerous modifications were made to the original algorithm, including structural changes to allow for seismic images as inputs, a change in the way edges are constructed for the graph, and a change in the weighting calculation for each edge. Each of these modifications increased the accuracy of the method when applied to seismic data.

Initial results from applying the modified algorithm to both synthetic and field seismic images are extremely encouraging. Segmentation of a synthetic image accurately located the boundaries of a complex salt body, although several different segments were required. Segmentation of both 2D and 3D field seismic data was even more successful. Compared to an existing implementation of the Normalized Cuts algorithm from Shi and Malik (2000), the new method performs extremely well - it required only half a minute to segment the synthetic data image (which is too large for the NCIS implementation to handle on a single processor), and only one second to segment the smaller field data example, over 150 times faster than NCIS. An additional advantage is that the newer algorithm is able to operate on the entire image, rather than only within a certain windowed radius of a previously interpreted boundary. This approach has many advantages, not least of which is the opportunity to identify segments other than only salt bodies. Instead of a binary salt/no-salt determination, the ability to identify coherent sedimentary ``segments'' as well would be tremendously useful for constructing seismic velocity models.

While these results are promising, there are many potential improvements that remain to be explored. First, the fact that the salt body is in some cases divided into multiple segments is a situation that should be examined. One possibility is to modify the edge weight calculation to include the relative importance of the amplitude/intensity and distance factors; right now, both terms are weighted equally (see equation 6). Another option is to change either the shape or size of the stencil (Figure 5) used to build the graph. Yet another potential improvement is the incorporation of multiple seismic attributes - for example, dip and frequency attributes. This enhancement can be accomplished relatively simply by taking multiple volumes as inputs, and calculating edge weights using some weighted combination of the different attributes at each pixel.

Interpreting subsurface salt bodies in large, 3D seismic images is an incredibly complex and tedious task. With further improvement, however, an accurate, efficient automatic segmentation scheme such as this one has the potential to be an extremely useful and powerful tool for processing and interpreting seismic images.

A new method for more efficient seismic image segmentation