A variety of special processing techniques have been developed to suppress noise or enhance signal on seismic data. Many of these use the transform-filter-inverse transform methodology previously introduced. This section deals principally with the attenuation of coherent dipping noise or random noise. Many conventional processes such as deconvolution, DMO, stacking and migration also attenuate noise - these are considered elsewhere. A further advanced section deals with the generation and suppression of multiple reflections. Prestack coherent noise may be created from the acquisition system itself or more likely from interference from other vessels. The process should understand the characteristics of the noise type to be removed before selecting the appropriate tool.
In this method the data are transformed into the FK domain either shot by shot (prestack) or by groups of traces (post-stack). Noise is separated from signal and identified in the FK domain. An appropriate filter is designed, applied and the FK inverse transform is performed. For post-stack processing many processing systems use standard symmetrical dip (fan or pie-slice) filters which pass dips of (for example) m 2ms/trace (harsh), m 4ms/trace (medium) and m 8ms/trace (mild). For prestack filtering of shots asymmetrical filters should normally be used unless an NMO correction is applied prior to filtering. Often the user can choose to pass the region between the selected dips (estimate signal) or reject the region between the dips (estimate noise). The former mode is usually chosen for convenience but runs some risk of corrupting the passed signal due to artifacts in the transforms. In the latter mode a percentage of the noise estimate may then be subtracted from the unfiltered data to pass signal which has not been through the transform process. If the dip limits chosen are too harsh or sufficiently gentle slopes not chosen for the filters the output data will look over-filtered or wormy.
Transforming the data to the FK domain, raising the amplitude spectrum to an exponential power and performing the inverse transform is a method sometimes used to reduce the levels of random noise. The method is available in PROMAX and Landmarks Post-Stack processing system.
This is the commonest modern technique for attenuating random noise since it has few artifacts and can be run in 2D or 3D modes. An important feature is the addback of original signal which can be tailored by the processor to produce a section with a pleasing appearence. Small temporal (e.g. 10 traces) and spatial (e.g. 20ms) windows of input data are Fourier transformed to the FX domain. Deconvolution operators are designed in the lateral (X) dimension to predict the coherent parts of the signal. Subtracting the coherent parts will leave the incoherent parts i.e. random noise which can then be inverse transformed and subtracted from the signal. The next window would then be selected, ensuring some overlap with the previous window.
An alternative to FK dip filtering in which the filters are designed in the tau-p domain. Generally this method would be more expensive, but does have the advantage that a time varying filter can easily be designed.
Coherency filtering is a term usually applied to time-domain dip filtering which is often rather harsh in nature.
Trace mixing is a cheap, rapid and somewhat harsh method of enhancing horizontal signal at the expense of dipping signal or random noise. A lateral window of 3, 5 or 7 traces would often be chosen and the traces weighted by a symmetrical weighting function such as (1,3,1) or (1,3,7,3,1) biased towards the central trace.
KL FILTERING ?
Care must be taken to avoid dip filtering prior to migration to avoid attenuating diffraction tails which