Random noise attenuation using forward-backward linear prediction

Random noise attenuation (RANNA) is commonly implemented as linear predicti on filtering (on a par-frequency basis) in the spatial domain. This paper p resents an up-to-date version of the RANNA process, in which the filter is designed by forward-backward linear prediction with a transient-free formul ation. This new version of RANNA has two improvements over the conventional noise reduction filter FXDECON. One is to overcome the transient-error inh erent in the FXDECON, caused by zero-trace padding outside the available da ta gate. The other is to minimise prediction residues in both directions, i nstead of the forward prediction in FXDECON (which in practice flips over t he one-sided prediction operator to build a two-sided filter). The one dime nsional prediction filter for use on 2-D seismic data is also extended to t wo dimensional, for use on 3-D data. This two dimensional filter is designe d as an optimum all-azimuth prediction operator. The method has been tested on real data, both 2-D and 3-D seismic stacks.