Noise Suppression with Structure-Oriented Stacking

Seismic data are usually contaminated with noise, which, stated simply, refers to unwanted signal. Noise in seismic data can originate from various sources but processed seismic data may contain the following types of noise: random noise, steeply dipping coherent noise, aliased coherent noise that may appear to be random, and coherent multiples that are often subparallel to the reflectors of interest.

For seismic data loaded on workstations, an interpreter can suppress all but multiples by running various types of filters. The simplest is the mean filter, which represents the arithmetic running average of a given number of adjacent spatial samples, usually three for 2-D data and five for 3-D. Mean filters can be designed for their application along structure rather than in time-slice mode, generating a “structure-oriented” filter. However, mean filters smear lateral discontinuities in the seismic data and should be avoided. In contrast, a median filter uses the same samples but replaces the amplitude of the central sample with the median value of the amplitudes. Principal component filters go one step further by augmenting the five or more samples along structural dip and azimuth by 2K parallel five-sample slices above and below the target sample. Mathematically, the principal component generates a five-sample pattern that best represents the lateral variation in amplitude along the 2K+1 slices. In the absence of high-amplitude artifacts in the data in general, the principal component filter accurately preserves lateral changes in seismic amplitude and rejects noise. More details on the types of noise and their suppression can be found in a detailed article on this topic published in the October 2014 Geophysical Corner.

Image Caption

Figure 1: Vertical slices through a land seismic data volume through the original data displayed with (a) a blue-white-red color color bar, and (b) a gray scale color bar. The human eye sees edges better in monochrome, thereby enhancing the steeply dipping noise indicated by the yellow arrows. The same data after (c) structure-oriented PC filtering, and structure-oriented stacking using (d) three traces and (e) five traces. The rejected noise is obtained by simply (f) subtracting (d) from (b) and (g) subtracting (e) from (b). Note in (g) that the larger operator rejects more of the steeply dipping noise at the expense of increased smoothing in (e). Data courtesy of New Zealand Petroleum and Minerals.

Please log in to read the full article

Seismic data are usually contaminated with noise, which, stated simply, refers to unwanted signal. Noise in seismic data can originate from various sources but processed seismic data may contain the following types of noise: random noise, steeply dipping coherent noise, aliased coherent noise that may appear to be random, and coherent multiples that are often subparallel to the reflectors of interest.

For seismic data loaded on workstations, an interpreter can suppress all but multiples by running various types of filters. The simplest is the mean filter, which represents the arithmetic running average of a given number of adjacent spatial samples, usually three for 2-D data and five for 3-D. Mean filters can be designed for their application along structure rather than in time-slice mode, generating a “structure-oriented” filter. However, mean filters smear lateral discontinuities in the seismic data and should be avoided. In contrast, a median filter uses the same samples but replaces the amplitude of the central sample with the median value of the amplitudes. Principal component filters go one step further by augmenting the five or more samples along structural dip and azimuth by 2K parallel five-sample slices above and below the target sample. Mathematically, the principal component generates a five-sample pattern that best represents the lateral variation in amplitude along the 2K+1 slices. In the absence of high-amplitude artifacts in the data in general, the principal component filter accurately preserves lateral changes in seismic amplitude and rejects noise. More details on the types of noise and their suppression can be found in a detailed article on this topic published in the October 2014 Geophysical Corner.

Structure-Oriented Stacking

Another type of process that has been introduced recently in a popular interpretation and visualization software package is the structure-oriented stacking. Stacking of adjacent traces is another concept that can be utilized for noise attenuation, as stacking is considered as the biggest noise attenuator in seismic data processing. When applied along the dip and azimuth of seismic reflections, it is termed as structure-oriented stacking.

We show the application of a principal component structure-oriented filter to seismic data and its comparison with the structure-oriented stacking in figure 1. A vertical slice through the seismic amplitude volume is shown in figure 1a in the usual blue-white-red color display. Alistair Brown notes that the human eye better perceives edges in monochrome such as the simple gray scale color bar used in figure 1b. In this example, the edges are generated by steeply dipping migration operator aliasing artifacts (yellow arrows) that cut through the more gently dipping reflectors of interest. But, such dipping noise patterns may be seen clearly when the same data are displayed in variable density gray scale. Consequently, we exhibit the same data shown in figure 1a, now in gray scale in figure 1b. Because they are steeply dipping, these noise events exhibit low apparent frequency. After structure-oriented filtering, the random noise is suppressed but the edge preservation has preserved the steeply dipping coherent noise. Equivalent displays from the application of structure-oriented stacking using three and five traces are shown in figure 1d and e, respectively. The different plots obtained from subtracting the data from the input seismic data shown in figure 1a and b are shown in figure 1f and g. Notice, the three-trace application has reduced the dipping noise somewhat, but it is still present in figure 1d. The five-trace application (figure 1f) gets rid of the linear noise (figure 1g), but at the cost of a little more smoothing of the seismic data.

In figure 2 we show a similar set of displays from a land seismic dataset. Again, notice the dipping linear noise indicated with yellow block arrows is seen clearly on the gray seismic display in figure 2b. The three-trace structure-oriented stacking filters out the random noise and reduces the linear noise somewhat, as seen on the displays in figures 2c and e. However, the five-trace structure-oriented stacking filters out the random noise and does a good job reducing the linear noise, as seen on the displays in figures 2d and f, though at the expense of a little more smoothing of the seismic data.

In figure 3 we show a comparison of stratal slices through coherence volumes generated from the input seismic data (figure 3a) as well as the same data put through structure-oriented stacking using three traces (figure 3b) and five traces (figure 3c), respectively. We notice the suppression of low coherence noise seen on stratal slice in figure 3a is much reduced in the equivalent slice shown in figure 3b. However, it could be argued that perhaps the display in figure 3c looks overly smoothed.

Structure-oriented stacking will be a useful process to run for seismic data that have high levels of random noise. For seismic data with discontinuities and large fracture signatures, structure-oriented stacking would need to be run in edge-preserving mode.